All posts by AIhub

A screenshot from the new simulator that will be trialled for a special challenge at RoboCup2025.

The annual RoboCup event, where teams gather from across the globe to take part in competitions across a number of leagues, will this year take place in Brazil, from 15-21 July. In advance of kick-off, we spoke to two members of the RoboCup Soccer 3D Simulation League: Executive Committee Member Klaus Dorer, and Stefan Glaser, who is on the Maintenance Committee and who has been recently developing a new simulator for the League.

Could start by just giving us a quick introduction to the Simulation League?
Klaus Dorer: There are two Simulation Leagues in Soccer: the 2D Simulation League and the 3D Simulation League. The 2D Simulation League, as the name suggests, is a flat league where the players and ball are simulated with simplified physics and the main focus is on team strategy. The 3D Simulation League is much closer to real robots; it simulates 11 versus 11 Nao robots. The level of control is like with real robots, where you move each motor of the legs and the arms and so on to achieve movement.

I understand that you have been working on a new simulator for the 3D League. What was the idea behind this new simulator?
Klaus: The aim is to bring us closer to the hardware leagues so that the simulator can be more useful. The current simulator that we use in the 3D Simulation League is called SimSpark. It was created in the early 2000s with the aim of making it possible to play 11 vs 11 players. With the hardware constraints of that time, there had to be some compromises on the physics to be able to simulate 22 players at the same time. So the simulation is physically somewhat realistic, but not in the sense that it’s easy to transpose it to a real Nao robot.

Stefan Glaser: The idea for developing a new simulator has been around for a few years. SimSpark is a very powerful simulation framework. The base framework is domain independent (not soccer specific) and specific simulations are realized via plugins. It supports multiple physics engines in the backend and provides a flexible scripting interface for configuration and adaptations of the simulation. However, all this flexibility comes with the price of complexity. In addition to that, SimSpark uses custom robot model specifications and communication protocols, limiting the amount of available robot models and requiring teams to develop custom communication layers only for communicating with SimSpark. As a result of this, SimSpark has not been widely adopted in the RoboCup community.

With the new simulator, I would like to address these two major issues: complexity and standardization. In the ML community, the MuJoCo physics engine has become a very popular choice for learning environments after Google DeepMind acquired it and released it open source. Its standards for world and robot model specifications are widely adopted in the community and there exist a lot of ready-to-use robot model specifications for a wide variety of virtual as well as real-world robots. In the middle of last year, they (MuJoCo) added a feature which allows you to manipulate the world representation during simulation (adding and removing objects to / from the simulation while preserving the simulation state). This is one essential requirement we have in the simulation league, where we start with an empty field and then the agents connect on demand and form the teams. When this feature has been added, I decided to make a step forward and try to implement a new simulator for the 3D Simulation League based on MuJoCo. Initially, I wanted to start development in C/C++ to achieve maximum performance, but then decided to start in Python to reduce complexity and make it more accessible for other developers. I started development on Easter Monday so it’s not even three months old!

I think it might be useful to explain a little bit more about the setup of our league and the requirements of the simulator. If we take the FIFA game (on your favorite gaming device) as an example, there is one simulation happening which simulates 22 players and the decision making is part of the simulation having full access to the state of the world. In the 3D Simulation League we have two teams with 11 robots on the field, but we also have 22 individual agent softwares which are connected to the simulation server, each controlling one single robot. Each connected agent only receives sensor information related to their robot in the simulation. They are also only allowed to communicate via the server – there is no direct communication between the agents allowed in Simulation League. So we have a general setup where the simulation server has to be able to accept up to 22 connections and manage the situation there. This functionality has been the major focus for me for the last couple of months and this part is already working well. Teams can connect their agents, which will receive sensor information and can actuate joints of the robot in the simulation and so on. They are also able to select different robot models if they like.

An illustration of the simulator set-up.

Presumably the new simulator has a better representation of the physics of a real robot.
Klaus: Exactly. For example, how the motors are controlled is now a bit different and much closer to real robots. So when I did my first experiments, I saw the robot collapse and I thought it was exactly how a real robot would collapse! In SimSpark we also had falling robots but the motor control in the new simulator is different. Now you can control the motors by speed, by force, by position, which is much more flexible – it’s closer to what we know from real robots.

I think that, at least initially, it will be more difficult for the Simulation League teams to get the robots to do what they want them to do, because it’s more realistic. For example, in SimSpark the ground contact was much more forgiving. So if you step hard on the ground, you don’t fall immediately with a SimSpark robot but with a MuJoCo robot this will be much more realistic. Indeed, in real robots ground contact is somewhat less forgiving.

I had a question about the vision aspect – how do the individual agents “see” the position of the other agents on the field?
Stefan: We simulate a virtual vision pipeline on the server side. You have a restricted field of view of ±60° horizontally and vertically. Within that field of view you will detect the head, the arms, the feet of other players, or the ball, for example, or different features of the field. Similar to common real-world vision pipelines, each detection consists of a label, a direction vector and the distance information. The information has some noise on it like real robots have, too, but teams don’t need to process camera images. They get the detections directly from the simulation server.

We’ve previously had a discussion about moving towards getting camera images of the simulation to integrate into the vision pipeline on the agent side. This was never really realistic in SimSpark with the implementation we had there. However, it should be possible with MuJoCo. However, for the first version, I used the same way the traditional simulator treated the vision. This means that teams don’t need to train a vision model, and don’t need to handle camera images to get started. This reduces the load significantly and also shifts the focus of the problem towards motion and decision making.

Will the simulator be used at RoboCup 2025?
Stefan: We plan to have a challenge with a new simulator and I will try to provide some demo games. At the moment it’s not really in a state where you can play a whole competition.

Klaus: That’s usually how we proceed with new simulators. We would not move from one to the other without any intermediate step. We will have a challenge this year at RoboCup 2025 with the new MuJoCo simulator where each participating team will try to teach the robot to kick as far as possible. So, we will not be playing a whole game, we won’t have multiple robots, just a single robot stepping in front of the ball and kicking the ball. That’s the technical challenge for this year. Teams will get an idea of how the simulator works, and we’ll get an idea of what has to be changed in the simulator to proceed.

This new challenge will be voluntary, so we are not sure how many teams will participate. Our team (MagmaOffenburg) will certainly take part. It will be interesting to see how well the teams perform because no one knows how far a good kick is in this simulator. It’s a bit like in Formula One when the rules change and no one knows which team will be the leading team.

Do you have an idea of how much adaptation teams will have to make if and when you move to the new simulator for the full matches?
Stefan: As a long-term member of 3D Simulation League, I know the old simulator SimSpark pretty well, and know the protocols involved and how the processes work. So the first version of the new simulator is designed to use the same basic protocol, the same sensor information, and so on. The idea is that the teams can use the new simulator with minimal effort in adapting their current agent software. So they should be able to get started pretty fast.

Although, when designing a new platform, I would like to take the opportunity to make a step forward in terms of protocols, because I also want to integrate other Leagues in the long-term. They usually have other control mechanisms, and they don’t use the same protocol that is prominent in 3D Simulation. Therefore there has to be some flexibility in the future. But for the first version, the idea was to get the Simulation League ready with minimal effort.

Klaus: The big idea is that this is not just used in the 3D Simulation league, but also as a useful simulator for the Humanoid League and also for the Standard Platform League (SPL). So if that turns out to be true, then it will be completely successful. For the Kick Challenge this year, for example, we use a T1 robot that is a Humanoid League robot.

Could you say something about this simulation to real world (Sim2Real) aspect?
Stefan: We’d like it to be possible for the motions and behaviors in the simulator to be ported to real robots. From my point of view, it would be useful the other way round too.

We, as a Simulation League, usually develop for the Simulation League and therefore would like to get the behaviors running on a real robot. But the hardware teams usually have a similar issue when they want to test high-level decision making. They might have two to five robots on the field, and if they want to play a high-level decision-making match and train in that regard, they always have to deploy a lot of robots. If they also want to have an opponent, they have to double the amount of robots in order to play a game to see how the strategy would turn out. The Sim2Real aspect is also interesting for these teams, because they should be able to take what they deployed on the real robot and it should also work in the simulation. They can then use the simulation to train high-level skills like team play, player positioning and so on, which is a challenging aspect for the real robot leagues like SPL or the Humanoid Leagues.

Klaus: And the reason we know this is because we have a team in the Simulation League and we have a team in the Humanoid League. So that’s another reason why we are keen to bring these things closer together.

How does the refereeing work in the Simulation League?
Klaus: A nice thing about Simulation Leagues is that there is a program which knows the real state of the world so we can build in the referee inside the simulator and it will not fail. For things like offside, whether the ball passed the goal line, that’s fail safe. All the referee decisions are taken by the system itself. We have a human referee but they never need to intervene. However, there are situations where we would like artificial intelligence to play a role. This is not currently the case in SimSpark because the rules are all hard coded. We have a lot of fouls that are debatable. For example, there are many fouls that teams agree should not have been a foul, and other fouls that are not called that should have been. It would be a nice AI learning task to get some situations judged by human referees and then train an AI model to better determine the rules for what is a foul and what isn’t a foul. But this is currently not the case.

Stefan: On the new simulator I am not that far into the development that I have implemented the automatic referee yet. I have some basic set of rules which progress the game as such, but judging fouls and deciding on special situations is not yet implemented in the new simulator.

What are the next steps for developing the simulator?
Stefan: One of the next major steps will be to refine the physics simulation. For instance, even though there exists a ball in the simulation, it is not yet really well refined. There are a lot of physics parameters which we have to decide on to reflect the real world as good as possible. This will likely require a series of experiments in order to get to the correct values for various aspects. In this aspect I’m hoping for some engagement of the community, as it is a great research opportunity and I personally would prefer the community to decide on a commonly accepted parameter set based on a level of evidence that I can’t easily provide all by myself. So in case someone is interested in refining the physics of the simulation such that it best reflects the real world, you are welcome to join!

Another major next step will be the development of the automatic referee of the soccer simulation, deciding on fouls, handling misbehaving agents and so on. In the first version, foul conditions will likely be judged by an expert system specifically designed for this purpose. The simulation league has developed a set of foul condition specifications which I plan to adapt. In a second step, I would like to integrate and support the development of AI based foul detection models. But yeah, one step after the other.

What are you particularly looking forward to at RoboCup2025?
Klaus: Well, with our team we have been vice world champion seven times in a row. This year we are really hoping to make it to world champion. We are very experienced in getting losses in finals and this year we are looking forward to changing that, from a team perspective.

Stefan: I’m going to Brazil in order to promote the simulator, not just for the Simulation League, but also across the boundaries for the Humanoid Leagues and the SPL Leagues. I think that this simulator is a great chance to bring people from all the leagues together. I’m particularly interested in the specific requirements of all the teams of the different leagues. This understanding will help me tailor the new simulator towards their needs. This is one of my major highlights for this year, I would say.

You can find out more about the new simulator at the project webpage, and from the documentation.

Klaus Dorer is professor for artificial intelligence, autonomous systems and software engineering at Offenburg University, Germany. He is also a member of the Institute for Machine Learning and Analytics IMLA. He has been team leader of the RoboCup simulation league teams magmaFreiburg (since 1999), living systems, magmaFurtwangen and is now team leader of magmaOffenburg since 2009. Since 2014, he has also been part of the humanoid adult size league team Sweaty.

Stefan Glaser is teaching assistant for artificial intelligence and intelligent autonomous systems at the Offenburg University, Germany. He has been part of the RoboCup simulation league team magmaOffenburg since 2009 and the RoboCup humanoid adult size league team Sweaty since 2014.

Kick-off in a Small Size League match. Image credit: Nicolai Ommer.

RoboCup is an international scientific initiative with the goal of advancing the state of the art of intelligent robots, AI and automation. The annual RoboCup event is due to take place from 15-21 July in Salvador, Brazil. The Soccer component of RoboCup comprises a number of Leagues, with one of these being the Small Size League (SSL). We caught up with Executive Committee member Nicolai Ommer to find out more about the SSL, how the auto referees work, and how teams use AI.

Could start by giving us a quick introduction to the Small Size League?

In the Small Size League (SSL) we have 11 robots per team – the only physical RoboCup soccer league to have the full number of players. The robots are small, cylindrical robots on wheels and they can move in any direction. They are self-built by the teams, so teams have to do both the hardware and the programming, and a lot of things have to work together to make a team work. The AI is central. We don’t have agents, so teams have a central computer at the field where they can do all the computation and then they send the commands to the robots in different abstractions. Some teams will just send velocity commands, other teams send a target.

We have a central vision system – this is maintained by the League, and has been since 2010. There are cameras above the field to track all the robots and the ball, so everyone knows where the robots are.

The robots can move up to 4 meters per second (m/s), after this point it gets quite unstable for the robots. They can change direction very quickly, and the ball can be kicked at 6.5 m/s. It’s quite fast and we’ve already had to limit the kick speed. Previously we had a limit of 8 m/s and before that 10m/s. However, no robot can catch a ball with this speed, so we decided to reduce it and put more focus on passing. This gives the keeper and the defenders a chance to actually intercept a kick.

It’s so fast that for humans it’s quite difficult to understand all the things that are going on. And that’s why, some years ago, we introduced auto refs, which help a lot in tracking, especially things like collisions and so on, where the human referee can’t watch everything at the same time.

How do the auto refs work then, and is there more than one operating at the same time?

When we developed the current system, to keep things fair, we decided to have multiple implementations of an auto ref system. These independent systems implement the same rules and then we do a majority vote on the decisions.

To do this we needed a middle component, so some years ago I started this project to have a new game controller. This is the user interface (UI) for the human referee who sits at a computer. In the UI you see the current game state, you can manipulate the game state, and this component coordinates the auto refs. The auto refs can connect and report fouls. If only one auto ref detects the foul, it won’t count it. But, if both auto refs report the foul within the time window, then it is counted. Part of the challenge was to make this all visual for the operator to understand. The human referee has the last word and makes the final decision.

We managed to establish two implementations. The aim was to have three implementations, which makes it easier to form a majority. However, it still works with just two implementations and we’ve had this for multiple years now. The implementations are from two different teams who are still active.

How do the auto refs deal with collisions?

We can detect collisions from the data. However, even for human referees it’s quite hard to determine who was at fault when two robots collide. So we had to just define a rule, and all the implementations of the auto ref implement the same rule. We wrote in the rulebook really specifically how you calculate if a collision happened and who was at fault. The first consideration is based on the velocity – below 1.5m/s it’s not a collision, above 1.5m/s it is. There is also another factor, relating to the angle calculation, that we also take into account to determine which robot was at fault.

What else do the auto refs detect?

Other fouls include the kick speed, and then there’s fouls relating to the adherence to normal game procedure. For example, when the other team has a free kick, then the opposing robots should maintain a certain distance from the ball.

The auto refs also observe non-fouls, in other words game events. For example, when the ball leaves the field. That’s the most common event. This one is actually not so easy to detect, particularly if there is a chip kick (where the ball leaves the playing surface). With the camera lens, the parabola of the ball can make it look like it’s outside the field of play when it isn’t. You need a robust filter to deal with this.

Also, when the auto refs detect a goal, we don’t trust them completely. When a goal is detected, we call it a “possible goal”. The match is halted immediately, all the robots stop, and the human referee can check all the available data before awarding the goal.

You’ve been involved in the League for a number of years. How has the League and the performance of the robots evolved over that time?

My first RoboCup was in 2012. The introduction of the auto refs has made the play a lot more fluent. Before this, we also introduced the concept of ball placement, so the robots would place the ball themselves for a free kick, or kick off, for example.

From the hardware side, the main improvement in recent years has been dribbling the ball in one-on-one situations. There has also been an improvement in the specialized skills performed by robots with a ball. For example, some years ago, one team (ZJUNlict) developed robots that could pull the ball backwards with them, move around defenders and then shoot at the goal. This was an unexpected movement, which we hadn’t seen before. Before this you had to do a pass to trick the defenders. Our team, TIGERs Mannheim, has also improved in this area now. But it’s really difficult to do this and requires a lot of tuning. It really depends on the field, the carpet, which is not standardized. So there’s a little bit of luck that your specifically built hardware is actually performing well on the competition carpet.

The Small Size League Grand Final at RoboCup 2024 in Eindhoven, Netherlands. TIGERs Mannheim vs. ZJUNlict. Video credit: TIGERs Mannheim. You can find the TIGERs’ YouTube channel here.

What are some of the challenges in the League?

One big challenge, and also maybe it’s a good thing for the League, is that we have a lot of undergraduate students in the teams. These students tend to leave the teams after their Bachelor’s or Master’s degree, the team members all change quite regularly, and that means that it’s difficult to retain knowledge in the teams. It’s a challenge to keep the performance of the team; it’s even hard to reproduce what previous members achieved. That’s why we don’t have large steps forward, because teams have to repeat the same things when new members join. However, it’s good for the students because they really learn a lot from the experience.

We are continuously working on identifying things which we can make available for everyone. In 2010 the vision system was established. It was a huge factor, meaning that teams didn’t have to do computer vision. And we are currently looking at establishing standards for wireless communication – this is currently done by everyone on their own. We want to advance the League, but at the same time, we also want to have this nature of being able to learn, being able to do all the things themselves if they want to.

You really need to have a team of people from different areas – mechanical engineering, electronics, project management. You also have to get sponsors, and you have to promote your project, get interested students in your team.

Could you talk about some of the AI elements to the League?

Most of our software is script-based, but we apply machine learning for small, subtle problems.

In my team, for example, we do model calibration with quite simple algorithms. We have a specific model for the chip kick, and another for the robot. The wheel friction is quite complicated, so we come up with a model and then we collect the data and use machine learning to detect the parameters.

For the actual match strategy, one nice example is from the team CMDragons. One year you could really observe that they had trained their model so that, once they scored goal, they upvoted the strategy that they applied before that. You could really see that the opponent reacted the same way all the time. They were able to score multiple goals, using the same strategy again and again, because they learned that if one strategy worked, they could use it again.

For our team, the TIGERs, our software is very much based on calculating scores for how good a pass is, how well can a pass be intercepted, and how we can improve the situation with a particular pass. This is hard-coded sometimes, with some geometry-based calculations, but there is also some fine-tuning. If we score a goal then we track back and see where the pass came from and we give bonuses on some of the score calculations. It’s more complicated than this, of course, but in general it’s what we try to do by learning during the game.

People often ask why we don’t do more with AI, and I think the main challenge is that, compared to other use cases, we don’t have that much data. It’s hard to get the data. In our case we have real hardware and we cannot just do matches all day long for days on end – the robots would break, and they need to be supervised. During a competition, we only have about five to seven matches in total. In 2016, we started to record all the games with a machine-readable format. All the positions are encoded, along with the referee decisions, and everything is in a log file which we publish centrally. I hope that with this growing amount of data we can actually apply some machine learning algorithms to see what previous matches and previous strategies did, and maybe get some insights.

What plans do you have for your team, the TIGERs?

We have actually won the competition for the last four years. We hope that there will be some other teams who can challenge us. Our defence has not really been challenged so we have a hard time finding weaknesses. We actually play against ourselves in simulation.

One thing that we want to improve on is precision because there is still some manual work to get everything calibrated and working as precisely as we want it. If some small detail is not working, for example the dribbling, then it risks the whole tournament. So we are working on making all these calibration processes easier, and to do more automatic data processing to determine the best parameters. In recent years we’ve worked a lot on dribbling in the 1 vs 1 situations. This has been a really big improvement for us and we are still working on that.

About Nicolai

Nicolai Ommer is a Software Engineer and Architect at QAware in Munich, specializing in designing and building robust software systems. He holds a B.Sc. in Applied Computer Science and an M.Sc. in Autonomous Systems. Nicolai began his journey in robotics with Team TIGERs Mannheim, participating in his first RoboCup in 2012. His dedication led him to join the RoboCup Small Size League Technical Committee and, in 2023, the Executive Committee. Passionate about innovation and collaboration, Nicolai combines academic insight with practical experience to push the boundaries of intelligent systems and contribute to the global robotics and software engineering communities.

The RoboCupRescue arena at RoboCup2024, Eindhoven.

RoboCup is an international scientific initiative with the goal of advancing the state of the science of intelligent robots, AI and automation. The annual RoboCup event will take place from 15-21 July in Salvador, Brazil. The RoboCupRescue League is an important element of the competition and focuses on the challenges involved in search and rescue applications. We caught up with Adam Jacoff, co-founder of the RoboCupRescue league, former RoboCup Trustee, and chair of the organising committee, to find out more.

Could you start by giving us an overview of the Rescue League?

The RoboCupRescue League is now in its 25th year hosting competitions and workshops all around the world. We’re focused on developing autonomous robots that can enable emergency responders to perform extremely hazardous tasks from safer stand-off distances. That includes search and rescue scenarios in compromised or collapsed structures, for example, but there are many tasks that firefighters and other public safety organizations do every day that would be safer with robots doing the dangerous part. We are really the only League at RoboCup that’s focused on emergency responders, so our arenas appear a little more chaotic than the other leagues, but in a controlled sort of way. We have twenty standard test methods that we use as challenge tasks that emergency responders and our league of teams around the world have helped develop and validate over many years. These competitions do three things at once.

First, they guide the research with tangible terrains, obstacles, and tasks that are representative of emergency response operations. The researchers likely do not know any emergency responders, but we do, and we have distilled their requirements into twenty or so standard test methods. So the teams know that if they can solve the challenges presented to them in the test lanes, maybe not all twenty, but ten of them, or five of them, that combination of capabilities are applicable to some public safety mission. And the more autonomy the researchers can implement into their robots, the more likely the robots will be effective and easy to use in the field. The main objective of our league is to quicken the development of effective and reliable autonomous behaviors for these complex environments.

RoboCup is an excellent incubator for developing and evaluating cutting edge research and at the same time for developing, validating, and disseminating new standard test methods. Our standard test methods are reproducible and approach the reality of very complex and hazardous environments in an incremental way. We have increasingly difficult settings for each test. So we start easy in the Preliminaries with all the terrain lanes flat so teams can optimize their approaches for each particular challenge. By Semi-Finals the terrain lanes are inclined into crossover 15° slopes with a complex terrain obstacle in the middle. The teams start performing sequences of multiple lanes, so they need to intelligently switch between different behaviors to succeed. Then for the Finals we inject more complexity with slippery features for the robots to lose traction, pinch points to force more steering, and higher step-over obstacles to make the sequences even harder. This is how we guide and challenge both autonomous and remotely operated robots beyond their comfort zone.

Second, the intensity of the experience and education for engineers and computer scientists leads to recruiting opportunities and careers. Everybody understands that teams who do well at RoboCupRescue have done something special. They can go on to more intense research or can help implement that technology in the commercial market. Our League has inspired and validated new robotic approaches to very complex environments and helped make the jump from research to commercial implementation.

The third piece of the puzzle is our connection to various communities of emergency responders. Our sequence of test lanes is exactly what emergency responders use to support their procurement and training. They use trial data captured with the robot manufacturers to help decide what robot to buy. They also use the test lanes to train remote operators and measure their proficiency. So RoboCupRescue helps emergency responders understand what the next wave of emerging technologies can do. For example, the nuclear waste cleanup industry wants to credential their remote robot operators to ensure robots don’t get stranded within extremely harsh radiation-filled environments. The robots need to keep working for 1,000 hours or more to be cost effective. They also want to ensure remote operators are proficient enough to safely perform the intended tasks without crashing the robot on the stairs and blocking the way for all other robots. Or worse yet, force humans in protective suits to go in and clear the way. Different organizations come to RoboCupRescue competitions and see various aspects of the league in action that are useful to them. In this case, as an obstacle course to get a robot driver’s license.

At Eindhoven last year (for RoboCup2024) we had the local bomb squads provide what we call a “reverse demonstration.” All these researchers give demonstrations of their robots at their schools to sponsors, family, friends, and maybe even to emergency responders. But they never get to see what the emergency responders do for a living – how dangerous their jobs are. So the RoboCupRescue league typically pauses for a while during the competition and comes together to watch a bomb technician get into their padded suit and go through the sequence of test lanes doing all the tasks that we’re asking the robots to perform. The researchers see how much of a burden it is just to wear the padded suits, which are very heavy and require forced air into the helmet. It’s not only physically difficult, they’re walking to and interacting with known explosives to render them safe – which is unconscionable given the capabilities of robots these days. They need to be working remotely from a safe stand-off distance. These “reverse demonstrations” are likely the first time the researchers ever get to meet people in bomb suits, and understand that it is these emergency responders we are working to help. Whether it’s firefighters or police, a natural or manmade disaster, we’re trying to help emergency responders stay out of harm’s way. They need capable, reliable, and easy to use robots to deal with extreme hazards more safely and effectively.

The RoboCupRescue arena at ICRA2025 (one of the qualifying events for RoboCup2025). Credit: Adam Jacoff.

So you mentioned trying to promote autonomy over teleoperation. How do you go about trying to encourage people to use more autonomy in the league?

Competitions need to have simple and clear rules. They need to challenge robots, lead them toward practical solutions, and highlight teams that are successful in a variety of challenges. So RoboCupRescue has developed scoring tasks for mobility, dexterity, and mapping within complex terrains. To score mobility points, for example, the robot needs to drive from one end of a terrain lane to the other end, or ascend and descend obstacles like stairs. They earn 1 point for each successful traverse up to 10 total repetitions to add some statistical significance. The operator is always out of sight of the robot, so working only through their interface as if the robot is inside a building. The terrain lanes and obstacles start easy and get harder, but all teams are being evaluated in the same difficulty setting as the competition progresses, so the points stay the same.

If the robot can successfully traverse the lane or perform the dexterity tasks without the operator touching the interface, we call that autonomy. At the beginning of each traverse the operator can manually set a goal point at the other end of the lane. Then they hit “go” and we all watch and learn together. If the robot gets to the other end without the operator touching the controls, the team gets 4 points instead of 1 point for that autonomous traverse. That 4:1 one ratio appears to be enough incentive to make sure teams develop autonomous behaviors and try them at each start point. However, they can always repair back to remote operation anytime the robot stalls to score at the lesser rate, so the rules encourage trying autonomy first. Also, we give Best-In-Class Autonomy awards to recognize teams that earned the most points autonomously throughout the competition.

Why do we need autonomy in emergency response operations? First, because the onboard autonomy typically makes the robot easier to use for a remote operator, which can also mean it is more likely to succeed down range or maybe be more reliable. Also, inside buildings there are often radio communications drop-out zones. Think about your cell phone reception inside concrete buildings or in basements under buildings, these places still need to be searched even when radio communication from a remote stand-off location outside the structure doesn’t support real-time video or control. So each RoboCupRescue lane, and the entire Labyrinth or Maze used for mapping, is considered a “radio drop-out zone” with tasks that score 4 to 1 if you can perform them autonomously. That applies to mobility, dexterity, and mapping tasks.

An example of one of the lanes (challenges) that teams tackle during competition. Credit: Adam Jacoff.

Will there be any new innovations to look out for this year?

What you’re going to see new this year are four-legged robots, called quadrupeds, that now have wheels as feet. There was a commercial robot that came to RoboCup2024 in Eindhoven. I invited them to demonstrate in the test lanes while the teams were at lunch, and this new robot design literally dominated our Semi-Finals sequence. Now that was a demonstration, conducted with the operator in direct line-of-sight of the robot, so much easier than remotely controlling the robot in the same terrains. But it didn’t have a remote interface. That’s exactly what our league of teams do best by adding interfaces, sensors, mapping, manipulators, and the autonomy to make it all effective and easier to use remotely.

These legged robots with wheels are a literal step function improvement in mobility, the likes of which I’ve only seen twice in thirty plus years. They have a level of mobility relative to their size that is astonishing. So we’re turning that new robot into our league’s first “standard platform.”

One of the four-legged robots with wheels as feet. Image from video taken at ICRA 2025, of the robot tackling one of the test lanes. Credit: Adam Jacoff.

Yes, I wanted to ask about the League’s plans for introducing a standard platform. Could you talk more about that?

This robot will provide a common development platform across teams with inherently good mobility within our terrains and obstacles. The league is negotiating with the manufacturer so that teams can purchase it at a relatively low price, which will bring down the cost of entry for new teams into the league. That level of mobility within our complex terrains and obstacles makes the autonomy a lot easier to develop – teams can focus on higher-level planning because the lower-level autonomous gaits and behaviors are already relatively effective. Similarly equipped teams can collaborate more closely, compare performance more directly, and share software solutions.

As a league, we can also start introducing teaming tasks for multiple robots. What kind of teaming might we explore? Maybe a certain object of interest is too heavy to drag for one robot, but could three of them work together to get it done like a team of dogs pulling a sled? Or maybe to carry payloads and supplies down-range, or even a non-ambulatory victim in a so-called litter back up-range to the base of operations. There will be lots of possibilities.

This new class of robots will operate in all the same test lanes along with all the other wheeled, tracked, and other robots developed by teams. We score every robot similarly of course, but then can separately compare the quantitative trial results for different classes of robots. It might turn out that it’s unfair to compare a new legged robot with wheels with a slower but stronger tracked robot that’s been winning RoboCupRescue for years. It’s not like tracked robots don’t have their use cases in emergency response operations. As soon as you get to tasks that involve exerting forces on the environment, like turning valves, the robots benefit from being heavier and stronger, which typically means tracks with independent front and rear flippers to negotiate the complex terrains and climb stairs. Everyone doesn’t need to adopt the new robot to compete, teams should work on implementations that push the state of the science in whatever direction they can. Then our Best-In-Class Awards can recognize the most effective robots in a variety of like-kind implementations, wheeled, tracked, legged, and maybe now legged with wheels.

When humanoids come into the league, and they will at some point, they’ll be welcomed and will be compared directly against other humanoids at least initially. We can also aggregate performance across a class of robots to identify advantages and apparent weaknesses. Emergency responders can then look at all the trial results to see which systems might work in their environments.

Do you know how many teams will be using this new robot?

I think we have almost twenty teams involved this year, but I’m not sure how many will be using the new legged robot with wheels. I expect maybe three or four initially, which would make a reasonable bracket of teams that can collaborate and compare more closely than others. There is a team from Austria that was in the German Open that is planning to come to Brazil. I’m sure when more people see the robot in action they’ll want one. Especially any new team coming into the league starting from scratch. Meanwhile, the RoboCupRescue league is poised to evaluate that robot design fully, add essential operational capabilities, and push the state of the science a bit further.

Will there be any different challenges this year?

Yes, this year we’re adding a new version of our original Stepfield terrain. They were our first rather complex terrain that inspired drastic changes in robot designs twenty years ago. The Stepfield terrain ensures the robots are always in contact with multiple different terrain elevations, and there is no respite throughout, even for legged robots with wheels. For the past couple of years, we got away from the Stepfield terrain because it was the hardest terrain to replicate at a temporary venue. But the year off for COVID was a growth time for us in terms of making the test methods as inexpensive and stowable as possible, so that small research organizations can replicate the tests to practice, refine approaches, and evaluate themselves during their development process. We first found a way to fabricate new Stepfields terrains from purchased, stackable crates. Now we have a fabricated version that accomplishes the same objectives less expensively. So RoboCupRescue teams will validate that the test apparatus as fabricated can stand up to the constant rigor of a very wide variety of robots across hundreds of trials in less than a week. That’s why RoboCupRescue is an excellent incubator for standard test methods too.

Is there anything else you’d like to highlight about the League?

One of the key features of RoboCupRescue is that the test lanes fabricated for the competition typically become a robot test facility for a local organization, like a nearby emergency responder facility or university. It can be set up perpetually, or it can move periodically to support regional competitions within the country. We usually try to connect the local organizers with their nearest emergency responders who already have robots or are seeking to purchase robots. The Olympic Village is the best analogy because it stays behind after the competition is over to be useful to the local region. So as we host our 25th year of RoboCupRescue in Brazil, we add to our world map of standard test facilities and new friends that we can continue to collaborate with professionally, and hopefully help their regional emergency responders use robots to stay out of harm’s way.

About Adam

Adam Jacoff is a robotics research engineer at the National Institute of Standards and Technology (NIST) which is part of the U.S. Department of Commerce. Over the past thirty years he has developed a variety of innovative robots and directed evaluations of more than a hundred others in a range of sizes, including the first technology readiness level assessment of autonomous mobility for the U.S. Army’s Experimental Unmanned Vehicle (XUV) (2002-2003).   His current efforts are focused toward developing 50 standard test methods to objectively evaluate ground robots, aerial drones, and aquatic vehicles for public safety operations. He has conducted 50 international robot competitions using the test methods as challenge tasks to guide innovation and measure progress (2000-present), and more than a hundred robot exercises to refine and validate the test methods with emergency responders and robot manufacturers (2005-present). These include dozens of comprehensive robot evaluations using the test methods to quantify key capabilities guiding more than $200M of purchasing decisions for civilian and military organizations (2010-present). He is now validating use of the test methods as repeatable practice tasks to focus remote operator/pilot training and measure proficiency for credentialing (2015-present).  He received a B.S. degree in Mechanical Engineering from the University of Maryland and a M.S. degree in Computer Science from Johns Hopkins University.

Action from RoboCupJunior Rescue at RoboCup 2024. Photo: RoboCup/Bart van Overbeeke.

The annual RoboCup event, where teams gather from across the globe to take part in competitions across a number of leagues, will this year take place in Brazil, from 15-21 July. An important part of the week is RoboCupJunior, which is designed to introduce RoboCup to school children, and sees hundreds of kids taking part in a variety of challenges across different leagues. This year, the lead organizer for RoboCupJunior is Ana Patrícia Magalhães. We caught up with her to find out how the preparations are going, what to expect at this year’s competition, and how RoboCup inspires communities.

Could you tell us about RoboCupJunior and the plans you have for the competition this year?

RoboCup will take place from 15-21 July, in Salvador, Brazil. We expect to receive people from more than 40 countries, across the Junior and Major Leagues. We are preparing everything to accommodate all the students taking part in RoboCupJunior, who will participate in the Junior Leagues of Soccer, Rescue and OnStage. They are children and teenagers, so we have organized shuttles to take them from the hotels to the convention center. We’ve also prepared a handbook with recommendations about security, places they can visit, places to eat. The idea is to provide all the necessary support for them, because they are so young. We’re also organizing a welcome party for the Juniors so that they can experience a little bit of our culture. It will hopefully be a good experience for them.

The Juniors will be located on the first level of the mezzanine at the convention center. They will be separate from the Major Leagues, who will be on the ground floor. Of course, they’ll be able to visit the Major Leagues, and talk to the students and other competitors there, but it will be nice for them to have their own space. There will also be some parents and teachers with them, so we decided to use this special, dedicated space.

RoboCupJunior On Stage at RoboCup 2024. Photo: RoboCup/Bart van Overbeeke.

Do you have any idea of roughly how many teams will be taking part?

Yes, so we’ll have about 48 teams in the Soccer Leagues, 86 teams in the Rescue Leagues, and 27 in OnStage. That’s a lot of teams. Each team has about three or four students, and many of the parents, teachers and professors travel with them too. In total, we expect about 600 people to be associated with RoboCupJunior.

RoboCupJunior Soccer at RoboCup 2024. Photo: RoboCup/Bart van Overbeeke.

Have you got more RoboCupJunior participants from Brazil this year due to the location?

Yes, we have many teams from Brazil competing. I don’t know the exact number, but there are definitely more Brazilian teams this year, because it’s a lot cheaper and easier for them to travel here. When we have competitions in other countries, it’s expensive for them. For example, I have a team here in Salvador that qualified for the super regional event in the US and our team couldn’t go. They had qualified, but they couldn’t go because they didn’t have money to pay for the ticket. Now, it will be possible for all the Brazilian teams qualified to participate because it’s cheaper for them to come here. So it’s a big opportunity for development and to live the RoboCup experience. It’s very important for children and teenagers to share their research, meet people from other countries, and see what they are doing, and what research path they are following. They are very grateful for the opportunity to have their work tested against others. In a competition, it is possible to compare your research with others. So it’s different from conferences where you present a paper and show your work, but it’s not possible to compare and evaluate the results with other similar work. In a competition you have this opportunity. It’s a good way to get insights and improve your research.

RoboCupJunior Rescue at RoboCup 2024. Photo: RoboCup/Bart van Overbeeke.

Your role at this RoboCup will be organizing RoboCupJunior. Are you also involved in the Major Leagues?

Yes, so my main role is organizing RoboCupJunior and I am also one of the chairs of the RoboCup Symposium. Besides, some teams from my laboratory are competing in the Major leagues. My team participates in the @Home league, but I haven’t had much time to help them recently, with all the preparations for RoboCup2025. Our laboratory also has teams from the 3d Simulation Soccer League, and the Flying Robots Demo. This will be the first time we’ll see a flying robot demo league at a RoboCup.

We’ll also have two junior teams from the Rescue Simulation League. They are very excited about taking part.

RoboCupJunior Rescue at RoboCup 2024. Photo: RoboCup/Bart van Overbeeke.

RoboCup was last held in Brazil in 2014, and I understand that there were quite a lot of new people that were inspired to join a team after that. Do you think the 2025 RoboCup will have the same effect and will inspire more people in Brazil to take part?

Yes, I hope so. The last one inspired many, many students. We could perceive the difference before and after RoboCup at that time, related to projects in schools. In 2014, RoboCup was held in João Pessoa, a city in the north east that is not as developed or populated as many other states in Brazil. It really improved the research in that place and the interest in robotics especially. After the 2014 RoboCup, we’ve had many projects submitted to the Brazilian RoboCup competition from that state every year. We believe that it was because of RoboCup being held there.

We hope that RoboCup2025 next month will have the same effect. We think it might have an even bigger impact, because there is more social media now and the news can spread a lot further. We are expecting many visitors. We will have a form where schools that want to visit can enroll on a guided visit of RoboCup. This will go live on the website next week, but we are already receiving many messages from schools asking how they can participate with their group. They are interested in the events, so we have high expectations.

We have been working on organizing RoboCup2025 for over a year, and there is still much to do. We are excited to receive everybody here, both for the competition and to see the city. We have a beautiful city on the coast, and some beautiful places to visit, so I recommend that people come and stay for some days after the competition to get to know our city.

About Ana Patrícia

Ana Patrícia F. Magalhães Mascarenhas received her PhD in Computer Science from the Federal University of Bahia (2016) and Master in Mechatronics from the Federal University of Bahia (2007). She is currently an adjunct professor at the State University of Bahia (UNEB) at the Information Systems course. She is a researcher and vice coordinator of the Center for Research in Computer Architecture, Intelligent Systems and Robotics (ACSO). Her current research focuses on service robotics and software engineering, especially related to the use of Artificial Intelligence (AI) in the software development process and in Model-Driven Development (DDM).

The Salvador Convention Center, where RoboCup 2025 will take place.

RoboCup is an international scientific initiative with the goal of advancing the state of the art of intelligent robots, AI and automation. The annual RoboCup event, where teams gather from across the globe to take part in competitions across a number of leagues, will this year take place in Brazil, from 15-21 July. We spoke to Marco Simões, one of the General Chairs of RoboCup 2025 and President of RoboCup Brazil, to find out what plans they have for the event, some new initiatives, and how RoboCup has grown in Brazil over the past ten years.

Marco Simões

Could you give us a quick introduction to RoboCup 2025?

RoboCup will be held in Salvador, Brazil. When RoboCup was held in Brazil 11 years ago, in 2014, we had a total of 100,000 visitors, so that was a great success. This year, we expect even more, around 150,000, during all the events. Nowadays, AI and robotics are attracting more attention. We are also in a town, Salvador, with a bigger population than the previous location (João Pessoa). For these reasons, we estimate the attendance to be about 150,000 people.

Regarding the number of teams, registration has not closed yet, so we’re unsure about the final numbers. However, we expect to have about 300-400 teams and around 3000 competitors. We have been helping with visas, so we hope to see higher participation from teams who couldn’t attend in the previous two years due to visa issues. We are doing our best to ensure people can come and have fun at RoboCup!

This is also a great year for the RoboCup community: We have just agreed on new global league partners, including the Chinese companies Unitree, Fourier, and Booster Robotics. They will bring their humanoids and four-legged robots to RoboCup. These will not only be exhibited to the public but also used by some teams. They are amazing robots with very good skills. So, I think this will be an amazing edition of RoboCup this year.

Did the 2014 event in Brazil inspire more teams to participate in RoboCup?

Yes, we have seen a significant increase in our RoboCup community. In the last two years, Brazil has had the fourth-largest number of teams and participants at RoboCup in Bordeaux (2023) and Eindhoven (2024). This was a very big increase because ten years ago, we were not even in the top eight or nine.

We’ve made a significant effort with RoboCupJunior in the last ten years. Most people who’ve taken part in RoboCupJunior have carried on and joined the RoboCup Major League. So, the number of teams in Brazil has been increasing year by year over the last ten years. This year, we have a great number of participants because of the lower travel costs. We are expecting to be in the top three this year in terms of the highest number of participants.

Photo of participants at RoboCup 2024, which took place in Eindhoven. Photo credit: RoboCup/Bart van Overbeeke

It’s impressive that so many RoboCupJunior participants go on to join a Major League team.

Yes, we have an initiative here in Brazil called the Brazilian Robotics Olympiad. In this event, we chose two Junior Leagues – OnStage and the Rescue Line League – and we organized a competition based on these two Leagues. We run it in regional competitions all over Brazil – so 27 states. We organize at least one competition in each state during the year, and the best teams from each state come to the national competition together with the Major Leagues. We organize the Brazilian Olympiad to get RoboCupJunior to more students. This is how we’ve managed to increase participation in RoboCupJunior. Then, when students go to university, many of them continue to participate, but in the Major Leagues. So that’s a very successful strategy we’ve used in Brazil in the last 10 years.

Could you tell us about some more of the highlights from the Brazilian RoboCup community in recent years?

Two or three years ago, one of the Brazilian teams was the champion of RoboCup @Home. We have seen a big increase in the number of teams in the @Home League. In the national competition in Brazil, we have more than 12 teams participating. Back in 2014, we only had one team participating. So we’ve had a great increase—this League is one of the highlights in Brazil.

More teams are also participating in the Small Size League (part of the soccer League). Two years ago, one of the Brazilian teams was the champion of the division B of the Small Size League. So, over the last five years, we’ve seen some Brazilian teams in the top three positions in Major Leagues in the RoboCup world competition. This is a result of the increase in the number of teams and the quality of what the teams are developing. So at this time, we have an increased number of publications and teams participating in the competition with good results, so that’s very important.

Another excellent contribution for this year is a league we created five years ago – a flying robot league, where autonomous drones perform some missions and tasks. We’ve proposed this League as a demo for RoboCup2025, and we will have a Flying Robot Demo at the competition this year. This will be the first time we’ll have autonomous drones at the RoboCup competition, and the Brazilian community proposed it.

RoboCup @Home with Toyota HSR robots in the Domestic Standard Platform League, RoboCup 2024. Photo: RoboCup/Bart van Overbeeke.

Will you be taking part in the competition this year, or will you be focusing entirely on your role as General Chair?

This year, my laboratory (ACSO/Uneb) has qualified for the 3d Simulation League (soccer), the Flying Robot Demo, and RoboCup @Home, so we are participating in three Leagues. We also supervise RoboCupJunior Teams in the Rescue Simulation League. This year, my students have had only a little supervision from me because I’ve been very engaged with the organization.

In our 3D simulation team, we have lots of developments with deep reinforcement learning and some new novel strategies that allow our teams to gain new skills, and we are combining the new skills with our former multi-agent coordination strategy. For this reason, I think we will have a robust team in the competition because we are not only working on skills, we are also working on multi-agency strategies. When both aspects are joined, you can have a really good soccer team that plays very well. We have a good team and expect to achieve a greater position this year. In the latter years, we were in the top four or five, but we hope to get into the top three this year.

In 3D, you not only work on multi-agent aspects but also need to work on skills such as walking, kicking, and running. Teams are now trying to develop new skills. For example, in recent years, our team has developed the sprint running movement, which was a result of deep reinforcement learning. It is not a natural running motion but a running movement that works according to the League’s rules. It makes the robots go very fast from one point to another, making the team very competitive.

Most teams are learning skills but don’t know how to exploit them strategically in the game. Our focus is not only on creating new skills but also on using them strategically. We are currently working on a very innovative approach.

This year, the simulation league will run a challenge using a new simulator based on MuJoCo. If the challenge goes well, we may move to this new simulator in the following years, which can more realistically simulate real humanoid robots.

Action from the semi-finals of RoboCup Soccer Humanoid League at RoboCup 2024. Photo: RoboCup/Bart van Overbeeke.

Finally, is there anything else you’d like to highlight about RoboCup2025?

We are working on partnerships with local companies. For example, we have sponsorship from Petrobras, one of the biggest oil companies in the world. They will discuss how they are using robotics and AI in their industry. They were also one of the first sponsors of the Flying Robots League. It’s important to have these links between industry and the RoboCup community.

We also have excellent support from local companies and the government. They will be showing the community their latest developments. In the Rescue League, for example, we’ll have a demonstration from the local force showing what they do to support people in disaster situations.

This event is also an excellent opportunity for RoboCuppers, especially those who have never been to Brazil, to spend some days after the event in Salvador, visiting some tourist spots. Salvador was the first Brazilian capital, so we have a rich history. There are a lot of historical sites to see and some great entertainment options, such as beaches or parties. People can have fun and enjoy the country!

About Marco

Marco Simões is an Associate Professor at Bahia State University, Salvador, Brazil. He is the General Chair of RoboCup2025, and President of RoboCup Brazil.

In this interview series, we’re meeting some of the AAAI/SIGAI Doctoral Consortium participants to find out more about their research. The Doctoral Consortium provides an opportunity for a group of PhD students to discuss and explore their research interests and career objectives in an interdisciplinary workshop together with a panel of established researchers. In this latest interview, we hear from Amar Halilovic, a PhD student at Ulm University.

Tell us a bit about your PhD – where are you studying, and what is the topic of your research?

I’m currently a PhD student at Ulm University in Germany, where I focus on explainable AI for robotics. My research investigates how robots can generate explanations of their actions in a way that aligns with human preferences and expectations, particularly in navigation tasks.

Could you give us an overview of the research you’ve carried out so far during your PhD?

So far, I’ve developed a framework for environmental explanations of robot actions and decisions, especially when things go wrong. I have explored black-box and generative approaches for the generation of textual and visual explanations. Furthermore, I have been working on planning of different explanation attributes, such as timing, representation, duration, etc. Lately, I’ve been working on methods for dynamically selecting the best explanation strategy depending on the context and user preferences.

Is there an aspect of your research that has been particularly interesting?

Yes, I find it fascinating how people interpret robot behavior differently depending on the urgency or failure context. It’s been especially rewarding to study how explanation expectations shift in different situations and how we can tailor explanation timing and content accordingly.

What are your plans for building on your research so far during the PhD – what aspects will you be investigating next?

Next, I’ll be extending the framework to incorporate real-time adaptation, enabling robots to learn from user feedback and adjust their explanations on the fly. I’m also planning more user studies to validate the effectiveness of these explanations in real-world human-robot interaction settings.

Amar with his poster at the AAAI/SIGAI Doctoral Consortium at AAAI 2025.

What made you want to study AI, and, in particular, explainable robot navigation?

I’ve always been interested in the intersection of humans and machines. During my studies, I realized that making AI systems understandable isn’t just a technical challenge—it’s key to trust and usability. Robot navigation struck me as a particularly compelling area because decisions are spatial and visual, making explanations both challenging and impactful.

What advice would you give to someone thinking of doing a PhD in the field?

Pick a topic that genuinely excites you—you’ll be living with it for several years! Also, build a support network of mentors and peers. It’s easy to get lost in the technical work, but collaboration and feedback are vital.

Could you tell us an interesting (non-AI related) fact about you?

I have lived and studied in four different countries.

About Amar

Amar is a PhD student at the Institute of Artificial Intelligence of Ulm University in Germany. His research focuses on Explainable Artificial Intelligence (XAI) in Human-Robot Interaction (HRI), particularly how robots can generate context-sensitive explanations for navigation decisions. He combines symbolic planning and machine learning to build explainable robot systems that adapt to human preferences and different contexts. Before starting his PhD, he studied Electrical Engineering at the University of Sarajevo in Sarajevo, Bosnia and Herzegovina, and Computer Science at Mälardalen University in Västerås, Sweden. Outside academia, Amar enjoys travelling, photography, and exploring connections between technology and society.

The AAMAS 2025 best paper and demo awards were presented at the 24th International Conference on Autonomous Agents and Multiagent Systems, which took place from 19-23 May 2025 in Detroit. The Distinguished Dissertation Award was also recently announced. The winners in the various categories are as follows:

Best Paper Award

Winner

• Soft Condorcet Optimization for Ranking of General Agents, Marc Lanctot, Kate Larson, Michael Kaisers, Quentin Berthet, Ian Gemp, Manfred Diaz, Roberto-Rafael Maura-Rivero, Yoram Bachrach, Anna Koop, Doina Precup

Finalists

• Azorus: Commitments over Protocols for BDI Agents, Amit K. Chopra, Matteo Baldoni, Samuel H. Christie V, Munindar P. Singh
• Curiosity-Driven Partner Selection Accelerates Convention Emergence in Language Games, Chin-Wing Leung, Paolo Turrini, Ann Nowe
• Reinforcement Learning-based Approach for Vehicle-to-Building Charging with Heterogeneous Agents and Long Term Rewards, Fangqi Liu, Rishav Sen, Jose Paolo Talusan, Ava Pettet, Aaron Kandel, Yoshinori Suzue, Ayan Mukhopadhyay, Abhishek Dubey
• Ready, Bid, Go! On-Demand Delivery Using Fleets of Drones with Unknown, Heterogeneous Energy Storage Constraints, Mohamed S. Talamali, Genki Miyauchi, Thomas Watteyne, Micael Santos Couceiro, Roderich Gross

Pragnesh Jay Modi Best Student Paper Award

Winners

• Decentralized Planning Using Probabilistic Hyperproperties, Francesco Pontiggia, Filip Macák, Roman Andriushchenko, Michele Chiari, Milan Ceska
• Large Language Models for Virtual Human Gesture Selection, Parisa Ghanad Torshizi, Laura B. Hensel, Ari Shapiro, Stacy Marsella

Runner-up

• ReSCOM: Reward-Shaped Curriculum for Efficient Multi-Agent Communication Learning, Xinghai Wei, Tingting Yuan, Jie Yuan, Dongxiao Liu, Xiaoming Fu

Finalists

• Explaining Facial Expression Recognition, Sanjeev Nahulanthran, Leimin Tian, Dana Kulic, Mor Vered
• Agent-Based Analysis of Green Disclosure Policies and Their Market-Wide Impact on Firm Behavior, Lingxiao Zhao, Maria Polukarov, Carmine Ventre

Blue Sky Ideas Track Best Paper Award

Winner

• Grounding Agent Reasoning in Image Schemas: A Neurosymbolic Approach to Embodied Cognition, François Olivier, Zied Bouraoui

Finalist

• Towards Foundation-model-based multiagent system to Accelerate AI for social impact, Yunfan Zhao, Niclas Boehmer, Aparna Taneja, Milind Tambe

Best Demo Award

Winner

• Serious Games for Ethical Preference Elicitation, Jayati Deshmukh, Zijie Liang, Vahid Yazdanpanah, Sebastian Stein, Sarvapali Ramchurn

Victor Lesser Distinguished Dissertation Award

The Victor Lesser Distinguished Dissertation Award is given for dissertations in the field of autonomous agents and multiagent systems that show originality, depth, impact, as well as quality of writing, supported by high-quality publications.

Winner

• Jannik Peters. Thesis title: Facets of Proportionality: Selecting Committees, Budgets, and Clusters

Runner-up

• Lily Xu. Thesis title: High-stakes decisions from low-quality data: AI decision-making for planetary health

The 2025 IEEE International Conference on Robotics and Automation (ICRA) best paper winners and finalists in the various different categories have been announced. The recipients were revealed during an award ceremony at the conference, which took place from 19-23 May in Atlanta, USA.

IEEE ICRA Best Paper Award on Robot Learning

Winner

• *Robo-DM: Data Management for Large Robot Datasets, Kaiyuan Chen, Letian Fu, David Huang, Yanxiang Zhang, Yunliang Lawrence Chen, Huang Huang, Kush Hari, Ashwin Balakrishna, Ted Xiao, Pannag Sanketi, John Kubiatowicz, Ken Goldberg

Finalists

• Achieving Human Level Competitive Robot Table Tennis, David D’Ambrosio, Saminda Wishwajith Abeyruwan, Laura Graesser, Atil Iscen, Heni Ben Amor, Alex Bewley, Barney J. Reed, Krista Reymann, Leila Takayama, Yuval Tassa, Krysztof Choromanski, Erwin Coumans, Deepali Jain, Navdeep Jaitly, Natasha Jaques, Satoshi Kataoka, Yuheng Kuang, Nevena Lazic, Reza, Mahjourian, Sherry Moore, Kenneth Oslund, Anish Shankar, Vikas Sindhwani, Vincent Vanhoucke, Grace Vesom, Peng Xu, Pannag Sanketi
• *No Plan but Everything under Control: Robustly Solving Sequential Tasks with Dynamically Composed Gradient Descent, Vito Mengers, Oliver Brock

IEEE ICRA Best Paper Award in Field and Service Robotics

Winner

• *PolyTouch: A Robust Multi-Modal Tactile Sensor for Contact-Rich Manipulation Using Tactile-Diffusion Policies, Jialiang Zhao, Naveen Kuppuswamy, Siyuan Feng, Benjamin Burchfiel, Edward Adelson

Finalists

• A New Stereo Fisheye Event Camera for Fast Drone Detection and Tracking, Daniel Rodrigues Da Costa, Maxime Robic, Pascal Vasseur, Fabio Morbidi
• *Learning-Based Adaptive Navigation for Scalar Field Mapping and Feature Tracking, Jose Fuentes, Paulo Padrao, Abdullah Al Redwan Newaz, Leonardo Bobadilla

IEEE ICRA Best Paper Award on Human-Robot Interaction

Winner

• *Human-Agent Joint Learning for Efficient Robot Manipulation Skill Acquisition, Shengchent Luo, Quanuan Peng, Jun Lv, Kaiwen Hong, Katherin Driggs-Campbell, Cewu Lu, Yong-Lu Li

Finalists

• *To Ask or Not to Ask: Human-In-The-Loop Contextual Bandits with Applications in Robot-Assisted Feeding, Rohan Banerjee, Rajat Kumar Jenamani, Sidharth Vasudev, Amal Nanavati, Katherine Dimitropoulou, Sarah Dean, Tapomayukh Bhattacharjee
• *Point and Go: Intuitive Reference Frame Reallocation in Mode Switching for Assistive Robotics, Allie Wang, Chen Jiang, Michael Przystupa, Justin Valentine, Martin Jagersand

IEEE ICRA Best Paper Award on Mechanisms and Design

Winner

• Individual and Collective Behaviors in Soft Robot Worms Inspired by Living Worm Blobs, Carina Kaeser, Junghan Kwon, Elio Challita, Harry Tuazon, Robert Wood, Saad Bhamla, Justin Werfel

Finalists

• *Informed Repurposing of Quadruped Legs for New Tasks, Fuchen Chen, Daniel Aukes
• *Intelligent Self-Healing Artificial Muscle: Mechanisms for Damage Detection and Autonomous, Ethan Krings, Patrick Mcmanigal, Eric Markvicka

IEEE ICRA Best Paper Award on Planning and Control

Winner

• *No Plan but Everything under Control: Robustly Solving Sequential Tasks with Dynamically Composed Gradient Descent, Vito Mengers, Oliver Brock

Finalists

• *SELP: Generating Safe and Efficient Task Plans for Robot Agents with Large Language Models, Yi Wu, Zikang Xiong, Yiran Hu, Shreyash Sridhar Iyengar, Nan Jiang, Aniket Bera, Lin Tan, Suresh Jagannathan
• *Marginalizing and Conditioning Gaussians Onto Linear Approximations of Smooth Manifolds with Applications in Robotics, Zi Cong Guo, James Richard Forbes, Timothy Barfoot

IEEE ICRA Best Paper Award in Robot Perception

Winner

• *MAC-VO: Metrics-Aware Covariance for Learning-Based Stereo Visual Odometry, Yuheng Qju, Yutian Chen, Zihao Zhang, Wenshan Wang, Sebastian Scherer

Finalists

• *Ground-Optimized 4D Radar-Inertial Odometry Via Continuous Velocity Integration Using Gaussian Process, Wooseong Yang, Hyesu Jang, Ayoung Kim
• *UAD: Unsupervised Affordance Distillation for Generalization in Robotic Manipulation, Yihe Tang, Wenlong Huang, Yingke Wang, Chengshu Li, Roy Yuan, Ruohan Zhang, Jiajun Wu, Li Fei-Fei

IEEE ICRA Best Paper Award in Robot Manipulation and Locomotion

Winner

• *D(R, O) Grasp: A Unified Representation of Robot and Object Interaction for Cross-Embodiment Dexterous Grasping, Zhenyu Wei, Zhixuan Xu, Jingxiang Guo, Yiwen Hou, Chongkai Gao, Zhehao Cai, Jiayu Luo, Lin Shao

Finalists

• *Full-Order Sampling-Based MPC for Torque-Level Locomotion Control Via Diffusion-Style Annealing, Haoru Xue, Chaoyi Pan, Zeji Yi, Guannan Qu, Guanya Shi
• *TrofyBot: A Transformable Rolling and Flying Robot with High Energy Efficiency, Mingwei Lai, Yugian Ye, Hanyu Wu, Chice Xuan, Ruibin Zhang, Qiuyu Ren, Chao Xu, Fei Gao, Yanjun Cao

IEEE ICRA Best Paper Award in Automation

Winner

• *Physics-Aware Robotic Palletization with Online Masking Inference, Tiangi Zhang, Zheng Wu, Yuxin Chen, Yixiao Wang, Boyuan Liang, Scott Moura, Masayoshi Tomizuka, Mingyu Ding, Wei Zhan

Finalists

• *In-Plane Manipulation of Soft Micro-Fiber with Ultrasonic Transducer Array and Microscope, Jieyun Zou, Siyuan An, Mingyue Wang, Jiaqi Li, Yalin Shi, You-Fu Li, Song Liu
• *A Complete and Bounded-Suboptimal Algorithm for a Moving Target Traveling Salesman Problem with Obstacles in 3D, Anoop Bhat, Geordan Gutow, Bhaskhar Vundurthy, Zhonggiang, Sivakumar Rathinam, Howie Choset

IEEE ICRA Best Paper Award in Medical Robotics

Winner

• *In-Vivo Tendon-Driven Rodent Ankle Exoskeleton System for Sensorimotor Rehabilitation, Juwan Han, Seunghyeon Park, Keehon Kim

Finalists

• *Image-Based Compliance Control for Robotic Steering of a Ferromagnetic Guidewire, An Hu, Chen Sun, Adam Dmytriw, Nan Xiao, Yu Sun
• *AutoPeel: Adhesion-Aware Safe Peeling Trajectory Optimization for Robotic Wound Care, Xiao Liang, Youcheng Zhang, Fei Liu, Florian Richter, Michael C. Yip

IEEE ICRA Best Paper Award on Multi-Robot Systems

Winner

• *Deploying Ten Thousand Robots: Scalable Imitation Learning for Lifelong Multi-Agent Path Finding, He Jiang, Yutong Wang, Rishi Veerapaneni, Tanishq Harish Duhan, Guillaume Adrien Sartoretti, Jiaoyang Li

Finalists

• Distributed Multi-Robot Source Seeking in Unknown Environments with Unknown Number of Sources, Lingpeng Chen, Siva Kailas, Srujan Deolasee, Wenhao Luo, Katia Sycara, Woojun Kim
• *Multi-Nonholonomic Robot Object Transportation with Obstacle Crossing Using a Deformable Sheet, Weijian Zhang, Charlie Street, Masoumeh Mansouri

IEEE ICRA Best Conference Paper Award

Winners

• *Marginalizing and Conditioning Gaussians Onto Linear Approximations of Smooth Manifolds with Applications in Robotics, Zi Cong Guo, James Richard Forbes, Timothy Barfoot
• *MAC-VO: Metrics-Aware Covariance for Learning-Based Stereo Visual Odometry, Yuheng Qju, Yutian Chen, Zihao Zhang, Wenshan Wang, Sebastian Scherer

In addition to the papers listed above, these paper were also finalists for the IEEE ICRA Best Conference Paper Award.

Finalists

• *MiniVLN: Efficient Vision-And-Language Navigation by Progressive Knowledge Distillation, Junyou Zhu, Yanyuan Qiao, Siqi Zhang, Xingjian He, Qi Wu, Jing Liu
• *RoboCrowd: Scaling Robot Data Collection through Crowdsourcing, Suvir Mirchandani, David D. Yuan, Kaylee Burns, Md Sazzad Islam, Zihao Zhao, Chelsea Finn, Dorsa Sadigh
• How Sound-Based Robot Communication Impacts Perceptions of Robotic Failure, Jai’La Lee Crider, Rhian Preston, Naomi T. Fitter
• *Obstacle-Avoidant Leader Following with a Quadruped Robot, Carmen Scheidemann, Lennart Werner, Victor Reijgwart, Andrei Cramariuc, Joris Chomarat, Jia-Ruei Chiu, Roland Siegwart, Marco Hutter
• *Dynamic Tube MPC: Learning Error Dynamics with Massively Parallel Simulation for Robust Safety in Practice, William Compton, Noel Csomay-Shanklin, Cole Johnson, Aaron Ames
• *Bat-VUFN: Bat-Inspired Visual-And-Ultrasound Fusion Network for Robust Perception in Adverse Conditions, Gyeongrok Lim, Jeong-ui Hong, Min Hyeon Bae
• *TinySense: A Lighter Weight and More Power-Efficient Avionics System for Flying Insect-Scale Robots, Zhitao Yu, Josh Tran, Claire Li, Aaron Weber, Yash P. Talwekar, Sawyer Fuller
• *TSCLIP: Robust CLIP Fine-Tuning for Worldwide Cross-Regional Traffic Sign Recognition, Guoyang Zhao, Fulong Ma, Weiging Qi, Chenguang Zhang, Yuxuan Liu, Ming Liu, Jun Ma
• *Geometric Design and Gait Co-Optimization for Soft Continuum Robots Swimming at Low and High Reynolds Numbers, Yanhao Yang, Ross Hatton
• *ShadowTac: Dense Measurement of Shear and Normal Deformation of a Tactile Membrane from Colored Shadows, Giuseppe Vitrani, Basile Pasquale, Michael Wiertlewski
• *Occlusion-aware 6D Pose Estimation with Depth-guided Graph Encoding and Cross-semantic Fusion for Robotic Grasping, Jingyang Liu, Zhenyu Lu, Lu Chen, Jing Yang, Chenguang Yang
• *Stable Tracking of Eye Gaze Direction During Ophthalmic Surgery, Tinghe Hong, Shenlin Cai, Boyang Li, Kai Huang
• *Configuration-Adaptive Visual Relative Localization for Spherical Modular Self-Reconfigurable Robots, Yuming Liu, Qiu Zheng, Yuxiao Tu, Yuan Gao, Guanqi Liang, Tin Lun Lam
• *Realm: Real-Time Line-Of-Sight Maintenance in Multi-Robot Navigation with Unknown Obstacles, Ruofei Bai, Shenghai Yuan, Kun Li, Hongliang Guo, Wei-Yun Yau, Lihua Xie

IEEE ICRA Best Student Paper Award

Winners

• *Deploying Ten Thousand Robots: Scalable Imitation Learning for Lifelong Multi-Agent Path Finding, He Jiang, Yutong Wang, Rishi Veerapaneni, Tanishq Harish Duhan, Guillaume Adrien Sartoretti, Jiaoyang Li
• *ShadowTac: Dense Measurement of Shear and Normal Deformation of a Tactile Membrane from Colored Shadows, Giuseppe Vitrani, Basile Pasquale, Michael Wiertlewski
• *Point and Go: Intuitive Reference Frame Reallocation in Mode Switching for Assistive Robotics, Allie Wang, Chen Jiang, Michael Przystupa, Justin Valentine, Martin Jagersand
• *TinySense: A Lighter Weight and More Power-Efficient Avionics System for Flying Insect-Scale Robots, Zhitao Yu, Josh Tran, Claire Li, Aaron Weber, Yash P. Talwekar, Sawyer Fuller

Note: papers with an * were eligible for the IEEE ICRA Best Student Paper Award.

The 2025 IEEE International Conference on Robotics & Automation (ICRA) took place from 19–23 May, in Atlanta, USA. The event featured plenary and keynote sessions, tutorial and workshops, forums, and a community day. Find out what the participants got up during the conference.

Check out what’s happening at the #ICRA2025 Welcome Reception! pic.twitter.com/w66IQDFsku

— IEEE ICRA (@ieee_ras_icra) May 19, 2025

The excitement is real — #ICRA2025 is already buzzing! pic.twitter.com/DtVgLwiaTB

— IEEE ICRA (@ieee_ras_icra) May 19, 2025

#ICRA #ICRA2025 #RoboticsInAfrica

[image or embed]

— Black in Robotics (@blackinrobotics.bsky.social) 18 May 2025 at 23:22

At #ICRA2025? Check out my student Yi Wu’s talk (TuCT1.4) at 3:30PM Tuesday in Room 302 at the Award Finalists 3 Session about how SELP Generates Safe and Efficient Plans for #Robot #Agents with #LLMs! #ConstrainedDecoding #LLMPlanner
@purduecs.bsky.social
@cerias.bsky.social

[image or embed]

— Lin Tan (@lin-tan.bsky.social) 19 May 2025 at 13:25

My MS student, Robel Mamo, is presenting his poster at #ICRA2025 #Field_Robotics workshop. His work is on "Crop-Aligned Cutout," a novel data augmentation method for under-canopy navigation @ieee_ras_icra @kennesawstate @KSUresearch pic.twitter.com/CBewagkpaQ

— Taeyeong Choi (최태영) (@ssuty) May 19, 2025

#ICRA2025 pic.twitter.com/FRfqgmSqNd

— Masato Kobayashi @るっと (@MeRTcooking) May 21, 2025

Malte Mosbach will present today 16:45 at #ICRA2025 in room 404 our paper:
"Prompt-responsive Object Retrieval with Memory-augmented Student-Teacher Learning"
www.ais.uni-bonn.de/videos/ICRA_…

[image or embed]

— Sven Behnke (@sven-behnke.bsky.social) 20 May 2025 at 15:57

#ICRA2025 pic.twitter.com/ANKoq3ry5K

— Masato Kobayashi @るっと (@MeRTcooking) May 21, 2025

Tomorrow morning at #ICRA2025, I will be presenting our findings on whether robots can learn dual-arm tasks from just a single demonstration. (Spoiler: they can!)

Come along!

This was led by my excellent PhD student Yilong Wang.

Paper & videos here: https://t.co/F2PtNJEJZT. pic.twitter.com/ycSAzuAgPn

— Edward Johns @ ICRA 2025 (@Ed__Johns) May 20, 2025

I will present our work on air-ground collaboration with SPOMP in 407A in a few minutes! We deployed 1 UAV and 3 UGVs in a fully autonomous mapping mission in large-scale environments. Come check it out! #ICRA2025 @grasplab.bsky.social

[image or embed]

— Fernando Cladera (@fcladera.bsky.social) 21 May 2025 at 20:13

Snapshots from #ICRA2025 @ieee_ras_icra : fans keep balloon walkers in constrained area @DennisHongRobot ; Artly coffee robot https://t.co/NCAMEo4fr5 ; mural ; and robo-friends #robots #artly #innovators pic.twitter.com/YAZdUWr9CD

— Heather Knight (@heatherknight) May 21, 2025

Cool things happening at #ICRA2025
RoboRacers gearing up for their qualifiers

[image or embed]

— Ameya Salvi (@ameyasalvi.bsky.social) 21 May 2025 at 13:56

Wednesday #ICRA2025 highlights included:

Plenary talk by Tessa Lau, CEO & Co-Founder, Dusty Robotics
Keynote & Technical Sessions
Community Day
ICRA Expo
Competitions
And more!

Check out tomorrow's events here: https://t.co/kS4WmAlZwJ pic.twitter.com/IboRH05KEI

— IEEE ICRA (@ieee_ras_icra) May 22, 2025

Our Community Building Day has been a success. LatinX in Robotics, Queer in Robotics, and Black in Robotics are just some of the incredible groups building community here at #ICRA2025! pic.twitter.com/EWNi0xY2NI

— IEEE ICRA (@ieee_ras_icra) May 21, 2025

New work at #ICRA2025!
Robust Robot Walker
We enable quadruped robots to pass tiny traps (bars, pits, poles) using only proprioception – no cameras, no depth!

Catch us at Thursday 16:55pm in Room 305!
https://t.co/571p4xTJ5c pic.twitter.com/03F2Gqf40D

— shaoting zhu (@ShaotingZ38103) May 22, 2025

Robot parade at @ieee_ras_icra #ICRA2025 pic.twitter.com/q2wHcQQN3R

— Sriram (@SriRam2528) May 22, 2025

Prof. Concha Monje presenting our BSc degree in Robotics Engineering at #ICRA2025 Forum on Undergraduate Robotics in Atlanta @uc3m @EPS_UC3M @ofic_eps_uc3m @ieeeras @ieee_ras_icra @mecanohumano https://t.co/R8fjbvJ6V3 pic.twitter.com/kDiWqSgx8M

— uc3mRoboticsLab (@uc3mRoboticsLab) May 22, 2025

We received the #ICRA2025 #HRI #Award for Arts & Robotics on our co-painting robot

This project shows how arts and robotics can be used as a testbed to create better robots and to discover new knowledge about humans.@ieee_ras_icra @UMRobotics @DARPA pic.twitter.com/T4VKxH2dKP

— patrícia alves-oliveira (@p_alvesoliveira) May 23, 2025

Fun times at the #ICRA2025 farewell reception! Celebrating Atlanta style with a block party. pic.twitter.com/rg71RPWWKc

— IEEE ICRA (@ieee_ras_icra) May 22, 2025

The 2025 IEEE International Conference on Robotics and Automation (ICRA) will take place from 19-23 May, in Atlanta, USA. The event will feature plenary talks, technical sessions, posters, workshops and tutorials, forums, and a science communication short course.

Plenary speakers

There are three plenary sessions this year. The speakers are as follows:

• Allison Okamura (Stanford University) – Rewired: The Interplay of Robots and Society
• Tessa Lau (Dusty Robotics) – So you want to build a robot company?
• Raffaello (Raff) D’Andrea (ETH Zurich) – Models are dead, long live models!

Keynote sessions

Tuesday 20, Wednesday 21 and Thursday 22 will see a total of 12 keynote sessions. The featured topics and speakers are:

• Rehabilitation & Physically Assistive Systems
• Brenna Argall
• Robert Gregg
• Keehoon Kim
• Christina Piazza
• Optimization & Control
• Todd Murphey
• Angela Schoellig
• Jana Tumova
• Ram Vasudevan
• Human Robot Interaction
• Sonia Chernova
• Dongheui Lee
• Harold Soh
• Holly Yanco
• Soft Robotics
• Robert Katzschmann
• Hugo Rodrigue
• Cynthia Sung
• Wenzhen Yuan
• Field Robotics
• Margarita Chli
• Tobias Fischer
• Joshua Mangelson
• Inna Sharf
• Bio-inspired Robotics
• Kyujin Cho
• Dario Floreano
• Talia Moore
• Yasemin Ozkan-Aydin
• Haptics
• Jeremy Brown
• Matej Hoffman
• Tania Morimoto
• Jee-Hwan Ryu
• Planning
• Hanna Kurniawati
• Jen Jen Chung
• Dan Halperin
• Jing Xiao
• Manipulation
• Tamim Asfour
• Yasuhisa Hasegawa
• Alberto Rodriguez
• Shuran Song
• Locomotion
• Sarah Bergbreiter
• Cosimo Della Santina
• Hae-Won Park
• Ludovic Righetti
• Safety & Formal Methods
• Chuchu Fan
• Meng Guo
• Changliu Liu
• Pian Yu
• Multi-robot Systems
• Sabine Hauert
• Dimitra Panagou
• Alyssa Pierson
• Fumin Zhang

Science communication training

Join Sabine Hauert, Evan Ackerman and Laura Bridgeman for a crash course on science communication. In this concise tutorial, you will learn how to share your work with a broader audience. This session will take place on 22 May, 11:00 – 12:15.

Workshops and tutorials

The programme of workshops and tutorials will take place on Monday 19 May and Friday 23 May. There are 59 events to choose from, and you can see the full list here.

Forums

There will be three forums as part of the programme, one each on Tuesday 20, Wednesday 21 and Thursday 22.

• Robot Ethics Forum
• Harnessing Learning, Data, Foundation Models and Open Source: How African Scientists are Advancing Robotics Research
• Undergraduate Robotics Education Programs: Structures, Platforms, and Approaches

Community building day

Wednesday 21 May is community building day, with six events planned:

• Queer in Robotics: Building a Community and Generating Inclusive Guidelines
• Harnessing Learning, Data, Foundation Models and Open Source: How African Scientists are Advancing Robotics Research
• RAS-WiE Voices — Women shaping the future of robotics and automation
• Black in Robotics and Blacks in Technology Social
• LatinX in AI & Robotics
• Community Building Day Dinner

Other events

You can find out more about the other sessions and event at the links below:

• Arts in robotics
• Career fair
• Competitions
• Stage presentations
• Technical tours

By Kristýna Janovská and Pavel Surynek

Imagine if all of our cars could drive themselves – autonomous driving is becoming possible, but to what extent? To get a vehicle somewhere by itself may not seem so tricky if the route is clear and well defined, but what if there are more cars, each trying to get to a different place? And what if we add pedestrians, animals and other unaccounted for elements? This problem has recently been increasingly studied, and already used in scenarios such as warehouse logistics, where a group of robots move boxes in a warehouse, each with its own goal, but all moving while making sure not to collide and making their routes – paths – as short as possible. But how to formalize such a problem? The answer is MAPF – multi-agent path finding [Silver, 2005].

Multi-agent path finding describes a problem where we have a group of agents – robots, vehicles or even people – who are each trying to get from their starting positions to their goal positions all at once without ever colliding (being in the same position at the same time).

Typically, this problem has been solved on graphs. Graphs are structures that are able to simplify an environment using its focal points and interconnections between them. These points are called vertices and can represent, for example, coordinates. They are connected by edges, which connect neighbouring vertices and represent distances between them.

If however we are trying to solve a real-life scenario, we strive to get as close to simulating reality as possible. Therefore, discrete representation (using a finite number of vertices) may not suffice. But how to search an environment that is continuous, that is, one where there is basically an infinite amount of vertices connected by edges of infinitely small sizes?

This is where something called sampling-based algorithms comes into play. Algorithms such as RRT* [Karaman and Frazzoli, 2011], which we used in our work, randomly select (sample) coordinates in our coordinate space and use them as vertices. The more points that are sampled, the more accurate the representation of the environment is. These vertices are connected to that of their nearest neighbours which minimizes the length of the path from the starting point to the newly sampled point. The path is a sequence of vertices, measured as a sum of the lengths of edges between them.

Figure 1: Two examples of paths connecting starting positions (blue) and goal positions (green) of three agents. Once an obstacle is present, agents plan smooth curved paths around it, successfully avoiding both the obstacle and each other.

We can get a close to optimal path this way, though there is still one problem. Paths created this way are still somewhat bumpy, as the transition between different segments of a path is sharp. If a vehicle was to take this path, it would probably have to turn itself at once when it reaches the end of a segment, as some robotic vacuum cleaners do when moving around. This slows the vehicle or a robot down significantly. A way we can solve this is to take these paths and smooth them, so that the transitions are no longer sharp, but smooth curves. This way, robots or vehicles moving on them can smoothly travel without ever stopping or slowing down significantly when in need of a turn.

Our paper [Janovská and Surynek, 2024] proposed a method for multi-agent path finding in continuous environments, where agents move on sets of smooth paths without colliding. Our algorithm is inspired by the Conflict Based Search (CBS) [Sharon et al., 2014]. Our extension into a continuous space called Continuous-Environment Conflict-Based Search (CE-CBS) works on two levels:

Figure 2: Comparison of paths found with discrete CBS algorithm on a 2D grid (left) and CE-CBS paths in a continuous version of the same environment. Three agents move from blue starting points to green goal points. These experiments are performed in the Robotic Agents Laboratory at Faculty of Information Technology of the Czech Technical University in Prague.

Firstly, each agent searches for a path individually. This is done with the RRT* algorithm as mentioned above. The resulting path is then smoothed using B-spline curves, polynomial piecewise curves applied to vertices of the path. This removes sharp turns and makes the path easier to traverse for a physical agent.

Individual paths are then sent to the higher level of the algorithm, in which paths are compared and conflicts are found. Conflict arises if two agents (which are represented as rigid circular bodies) overlap at any given time. If so, constraints are created to forbid one of the agents from passing through the conflicting space at a time interval during which it was previously present in that space. Both options which constrain one of the agents are tried – a tree of possible constraint settings and their solutions is constructed and expanded upon with each conflict found. When a new constraint is added, this information passes to all agents it concerns and their paths are re-planned so that they avoid the constrained time and space. Then the paths are checked again for validity, and this repeats until a conflict-free solution, which aims to be as short as possible is found.

This way, agents can effectively move without losing speed while turning and without colliding with each other. Although there are environments such as narrow hallways where slowing down or even stopping may be necessary for agents to safely pass, CE-CBS finds solutions in most environments.

This research is supported by the Czech Science Foundation, 22-31346S.

You can read our paper here.

References

• Janovská, K. and Surynek, P. (2024). Multi-agent Path Finding in Continuous Environment, CoRR.
• Sharon, G., Stern, R., Felner, A., and Sturtevant, N. R. (2014). Conflict-based search for optimal multi-agent pathfinding, Artificial Intelligence.
• Karaman, S. and Frazzoli, E. (2011). Sampling-based algorithms for optimal motion planning, CoRR.
• Piegl, L. and Tiller, W. (1996). The NURBS Book, Springer-Verlag, New York, USA, second edition.
• Silver, D. (2005). Cooperative pathfinding, Proceedings of the First Artificial Intelligence and Interactive Digital Entertainment Conference, Marina del Rey, California, USA.

Yuki Mitsufuji is a Lead Research Scientist at Sony AI. Yuki and his team presented two papers at the recent Conference on Neural Information Processing Systems (NeurIPS 2024). These works tackle different aspects of image generation and are entitled: GenWarp: Single Image to Novel Views with Semantic-Preserving Generative Warping and PaGoDA: Progressive Growing of a One-Step Generator from a Low-Resolution Diffusion Teacher . We caught up with Yuki to find out more about this research.

There are two pieces of research we’d like to ask you about today. Could we start with the GenWarp paper? Could you outline the problem that you were focused on in this work?

The problem we aimed to solve is called single-shot novel view synthesis, which is where you have one image and want to create another image of the same scene from a different camera angle. There has been a lot of work in this space, but a major challenge remains: when an image angle changes substantially, the image quality degrades significantly. We wanted to be able to generate a new image based on a single given image, as well as improve the quality, even in very challenging angle change settings.

How did you go about solving this problem – what was your methodology?

The existing works in this space tend to take advantage of monocular depth estimation, which means only a single image is used to estimate depth. This depth information enables us to change the angle and change the image according to that angle – we call it “warp.” Of course, there will be some occluded parts in the image, and there will be information missing from the original image on how to create the image from a new angle. Therefore, there is always a second phase where another module can interpolate the occluded region. Because of these two phases, in the existing work in this area, geometrical errors introduced in warping cannot be compensated for in the interpolation phase.

We solve this problem by fusing everything together. We don’t go for a two-phase approach, but do it all at once in a single diffusion model. To preserve the semantic meaning of the image, we created another neural network that can extract the semantic information from a given image as well as monocular depth information. We inject it using a cross-attention mechanism, into the main base diffusion model. Since the warping and interpolation were done in one model, and the occluded part can be reconstructed very well together with the semantic information injected from outside, we saw the overall quality improved. We saw improvements in image quality both subjectively and objectively, using metrics such as FID and PSNR.

Can people see some of the images created using GenWarp?

Yes, we actually have a demo, which consists of two parts. One shows the original image and the other shows the warped images from different angles.

Moving on to the PaGoDA paper, here you were addressing the high computational cost of diffusion models? How did you go about addressing that problem?

Diffusion models are very popular, but it’s well-known that they are very costly for training and inference. We address this issue by proposing PaGoDA, our model which addresses both training efficiency and inference efficiency.

It’s easy to talk about inference efficiency, which directly connects to the speed of generation. Diffusion usually takes a lot of iterative steps towards the final generated output – our goal was to skip these steps so that we could quickly generate an image in just one step. People call it “one-step generation” or “one-step diffusion.” It doesn’t always have to be one step; it could be two or three steps, for example, “few-step diffusion”. Basically, the target is to solve the bottleneck of diffusion, which is a time-consuming, multi-step iterative generation method.

In diffusion models, generating an output is typically a slow process, requiring many iterative steps to produce the final result. A key trend in advancing these models is training a “student model” that distills knowledge from a pre-trained diffusion model. This allows for faster generation—sometimes producing an image in just one step. These are often referred to as distilled diffusion models. Distillation means that, given a teacher (a diffusion model), we use this information to train another one-step efficient model. We call it distillation because we can distill the information from the original model, which has vast knowledge about generating good images.

However, both classic diffusion models and their distilled counterparts are usually tied to a fixed image resolution. This means that if we want a higher-resolution distilled diffusion model capable of one-step generation, we would need to retrain the diffusion model and then distill it again at the desired resolution.

This makes the entire pipeline of training and generation quite tedious. Each time a higher resolution is needed, we have to retrain the diffusion model from scratch and go through the distillation process again, adding significant complexity and time to the workflow.

The uniqueness of PaGoDA is that we train across different resolution models in one system, which allows it to achieve one-step generation, making the workflow much more efficient.

For example, if we want to distill a model for images of 128×128, we can do that. But if we want to do it for another scale, 256×256 let’s say, then we should have the teacher train on 256×256. If we want to extend it even more for higher resolutions, then we need to do this multiple times. This can be very costly, so to avoid this, we use the idea of progressive growing training, which has already been studied in the area of generative adversarial networks (GANs), but not so much in the diffusion space. The idea is, given the teacher diffusion model trained on 64×64, we can distill information and train a one-step model for any resolution. For many resolution cases we can get a state-of-the-art performance using PaGoDA.

Could you give a rough idea of the difference in computational cost between your method and standard diffusion models. What kind of saving do you make?

The idea is very simple – we just skip the iterative steps. It is highly dependent on the diffusion model you use, but a typical standard diffusion model in the past historically used about 1000 steps. And now, modern, well-optimized diffusion models require 79 steps. With our model that goes down to one step, we are looking at it about 80 times faster, in theory. Of course, it all depends on how you implement the system, and if there’s a parallelization mechanism on chips, people can exploit it.

Is there anything else you would like to add about either of the projects?

Ultimately, we want to achieve real-time generation, and not just have this generation be limited to images. Real-time sound generation is an area that we are looking at.

Also, as you can see in the animation demo of GenWarp, the images change rapidly, making it look like an animation. However, the demo was created with many images generated with costly diffusion models offline. If we could achieve high-speed generation, let’s say with PaGoDA, then theoretically, we could create images from any angle on the fly.

Find out more:

• GenWarp: Single Image to Novel Views with Semantic-Preserving Generative Warping, Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seungryong Kim, Yuki Mitsufuji.
• GenWarp demo
• PaGoDA: Progressive Growing of a One-Step Generator from a Low-Resolution Diffusion Teacher, Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Yuhta Takida, Naoki Murata, Toshimitsu Uesaka, Yuki Mitsufuji, Stefano Ermon.

About Yuki Mitsufuji

Yuki Mitsufuji is a Lead Research Scientist at Sony AI. In addition to his role at Sony AI, he is a Distinguished Engineer for Sony Group Corporation and the Head of Creative AI Lab for Sony R&D. Yuki holds a PhD in Information Science & Technology from the University of Tokyo. His groundbreaking work has made him a pioneer in foundational music and sound work, such as sound separation and other generative models that can be applied to music, sound, and other modalities.

In a series of interviews, we’re meeting some of the AAAI/SIGAI Doctoral Consortium participants to find out more about their research. In this latest interview, we hear from Amina Mević who is applying machine learning to semiconductor manufacturing. Find out more about her PhD research so far, what makes this field so interesting, and how she found the AAAI Doctoral Consortium experience.

Tell us a bit about your PhD – where are you studying, and what is the topic of your research?

I am currently pursuing my PhD at the University of Sarajevo, Faculty of Electrical Engineering, Department of Computer Science and Informatics. My research is being carried out in collaboration with Infineon Technologies Austria as part of the Important Project of Common European Interest (IPCEI) in Microelectronics. The topic of my research focuses on developing an explainable multi-output virtual metrology system based on machine learning to predict the physical properties of metal layers in semiconductor manufacturing.

Could you give us an overview of the research you’ve carried out so far during your PhD?

In the first year of my PhD, I worked on preprocessing complex manufacturing data and preparing a robust multi-output prediction setup for virtual metrology. I collaborated with industry experts to understand the process intricacies and validate the prediction models. I applied a projection-based selection algorithm (ProjSe), which aligned well with both domain knowledge and process physics.

In the second year, I developed an explanatory method, designed to identify the most relevant input features for multi-output predictions.

Is there an aspect of your research that has been particularly interesting?

For me, the most interesting aspect is the synergy between physics, mathematics, cutting-edge technology, psychology, and ethics. I’m working with data collected during a physical process—physical vapor deposition—using concepts from geometry and algebra, particularly projection operators and their algebra, which have roots in quantum mechanics, to enhance both the performance and interpretability of machine learning models. Collaborating closely with engineers in the semiconductor industry has also been eye-opening, especially seeing how explanations can directly support human decision-making in high-stakes environments. I feel truly honored to deepen my knowledge across these fields and to conduct this multidisciplinary research.

What are your plans for building on your research so far during the PhD – what aspects will you be investigating next?

I plan to focus more on time series data and develop explanatory methods for multivariate time series models. Additionally, I intend to investigate aspects of responsible AI within the semiconductor industry and ensure that the solutions proposed during my PhD align with the principles outlined in the EU AI Act.

How was the AAAI Doctoral Consortium, and the AAAI conference experience in general?

Attending the AAAI Doctoral Consortium was an amazing experience! It gave me the opportunity to present my research and receive valuable feedback from leading AI researchers. The networking aspect was equally rewarding—I had inspiring conversations with fellow PhD students and mentors from around the world. The main conference itself was energizing and diverse, with cutting-edge research presented across so many AI subfields. It definitely strengthened my motivation and gave me new ideas for the final phase of my PhD.

Amina presenting two posters at AAAI 2025.

What made you want to study AI?

After graduating in theoretical physics, I found that job opportunities—especially in physics research—were quite limited in my country. I began looking for roles where I could apply the mathematical knowledge and problem-solving skills I had developed during my studies. At the time, data science appeared to be an ideal and promising field. However, I soon realized that I missed the depth and purpose of fundamental research, which was often lacking in industry roles. That motivated me to pursue a PhD in AI, aiming to gain a deep, foundational understanding of the technology—one that can be applied meaningfully and used in service of humanity.

What advice would you give to someone thinking of doing a PhD in the field?

Stay curious and open to learning from different disciplines—especially mathematics, statistics, and domain knowledge. Make sure your research has a purpose that resonates with you personally, as that passion will help carry you through challenges. There will be moments when you’ll feel like giving up, but before making any decision, ask yourself: am I just tired? Sometimes, rest is the solution to many of our problems. Finally, find mentors and communities to share ideas with and stay inspired.

Could you tell us an interesting (non-AI related) fact about you?

I’m a huge science outreach enthusiast! I regularly volunteer with the Association for the Advancement of Science and Technology in Bosnia, where we run workshops and events to inspire kids and high school students to explore STEM—especially in underserved communities.

About Amina

Amina Mević is a PhD candidate and teaching assistant at the University of Sarajevo, Faculty of Electrical Engineering, Bosnia and Herzegovina. Her research is conducted in collaboration with Infineon Technologies Austria as part of the IPCEI in Microelectronics. She earned a master’s degree in theoretical physics and was awarded two Golden Badges of the University of Sarajevo for achieving a GPA higher than 9.5/10 during both her bachelor’s and master’s studies. Amina actively volunteers to promote STEM education among youth in Bosnia and Herzegovina and is dedicated to improving the research environment in her country.

Congratulations to Shlomo Zilberstein on winning the 2025 ACM/SIGAI Autonomous Agents Research Award. This prestigious award is made for excellence in research in the area of autonomous agents. It is intended to recognize researchers in autonomous agents whose current work is an important influence on the field.

Professor Shlomo Zilberstein was recognised for his work establishing the field of decentralized Markov Decision Processes (DEC-MDPs), laying the groundwork for decision-theoretic planning in multi-agent systems and multi-agent reinforcement learning (MARL). The selection committee noted that these contributions have become a cornerstone of multi-agent decision-making, influencing researchers and practitioners alike.

Shlomo Zilberstein is Professor of Computer Science and former Associate Dean of Research at the University of Massachusetts Amherst. He is a Fellow of AAAI and the ACM, and has received numerous awards, including the UMass Chancellor’s Medal, the IFAAMAS Influential Paper Award, and the AAAI Distinguished Service Award.

Image taken from the front cover of the Future of AI Research report.

The Association for the Advancement of Artificial Intelligence (AAAI), has published a report on the Future of AI Research. The report, which was announced by outgoing AAAI President Francesca Rossi during the AAAI 2025 conference, covers 17 different AI topics and aims to clearly identify the trajectory of AI research in a structured way.

The report is the result of a Presidential Panel, chaired by Francesca Rossi, and comprising of 24 experienced AI researchers, who worked on the project between summer 2024 and spring 2025. As well as the views of the panel members, the report also draws on community feedback, which was received from 475 AI researchers via a survey.

The 17 topics, each with a dedicated chapter, are as follows.

• AI Reasoning
• AI Factuality & Trustworthiness
• AI Agents
• AI Evaluation
• AI Ethics & Safety
• Embodied AI
• AI & Cognitive Science
• Hardware & AI
• AI for Social Good
• AI & Sustainability
• AI for Scientific Discovery
• Artificial General Intelligence (AGI)
• AI Perception vs. Reality
• Diversity of AI Research Approaches
• Research Beyond the AI Research Community
• Role of Academia
• Geopolitical Aspects & Implications of AI

Each chapter includes a list of main takeaways, context and history, current state and trends, research challenges, and community opinion. You can read the report in full here.