The C-suite loves what agentic AI promises: autonomous systems that can think, decide, and act without constant human intervention. The potential for productivity and lower costs is undeniable — until the bills start rolling in.
If your “strategy” is to ship first and figure out the cost later, you’re not building agentic AI. You’re financing a science project.
The goal is not to cut costs. It’s to engineer cost, speed, and quality to move together from day one. Because once an agent is in production, every weak decision you made in architecture, governance, and infrastructure becomes a recurring charge.
When cloud costs can spike by more than 200% overnight and development cycles stretch months beyond plan, that “transformative” agent stops looking like innovation and starts looking like a resource sink you can’t justify — to the board, to the business, or to your own team.
This isn’t another “how to save money on artificial intelligence” listicle. It reflects how leading teams using DataRobot align architecture, governance, and infrastructure with spend so autonomy doesn’t turn into a blank check. This is a comprehensive strategic framework for enterprise leaders who refuse to choose between innovation and financial discipline. We’ll surface the real cost drivers, call out where competitors routinely bleed money (so you don’t), and lay out infrastructure and operating strategies that keep your agentic AI initiatives from becoming cutting-room-floor casualties.
Agentic AI isn’t a reactive system that waits for inputs and spits out predictions. These are agents that act on their own, guided by the rules and logic you build into them. They’re contextually aware of their environment, learning from and making decisions by taking action across multiple connected systems, workflows, and business processes simultaneously.
That autonomy is the whole point — and it’s exactly why agentic AI gets expensive in a hurry.
The cost of autonomy hits you in three ways.
Where traditional AI might cost $0.001 per inference, agentic systems can run $0.10–$1.00 per complex decision cycle. Multiply that by hundreds or thousands of daily interactions, and you’re looking at monthly bills that are hard to defend, even when the use case is “working.”
An important component here is that hidden costs in agentic AI often dwarf the obvious ones. Compute costs aren’t the real budget killers. It’s the operational complexity that nobody talks about (until it’s too late).
Let’s cut through the vendor marketing and look at where your money actually goes. Agentic AI costs break down into four major buckets, each with its own optimization challenges and budget multipliers:
Enterprises getting this right aren’t necessarily spending less overall. They’re just a) using their dollars in smarter ways and b) understanding which categories offer the most optimization potential and cost controls for their architecture from day one.
The costs that ultimately kill agentic AI projects are the operational realities that show up only after your agents start making real decisions in production environments: real invoices, real headcount burn, and real executive scrutiny.
Your agentic AI system made 10,000 autonomous decisions overnight. Now, three customers are complaining about issues with their accounts. How do you debug that?
Traditional monitoring assumes you know what to look for. Agentic systems generate emergent behaviors that require entirely new observability approaches. You need to track decision paths, conversation flows, multi-agent interactions, tool calls, and the reasoning behind each action.
Here’s the expensive truth: Without proper observability, debugging turns into days of forensic work. That’s where labor costs quietly explode — engineers pulled off roadmap work, incident calls multiplying, and leadership demanding certainty you can’t provide because you didn’t instrument the system to explain itself.
Building observability into agent architecture is mandatory from the start. Selective logging, automated anomaly detection, and decision replay systems make debugging tractable without turning your platform into a logging furnace. And this is where unified platforms matter, because if your observability is stitched together across tools, your costs and blind spots multiply together, too.
Retrofitting governance and security controls onto autonomous systems that are already making production decisions can turn your “cheap” agentic AI implementation into an expensive rewrite.
A few requirements are non-negotiable for enterprise deployments:
Each adds another layer and cost that scales as your agent ecosystem grows.
The reality is that misbehaving AI costs scale with autonomy. When a traditional system makes a bad prediction, you can often catch it downstream. But when an agent takes incorrect actions across multiple business processes, damage branches fast, and you pay twice: once to fix the problem and again to restore trust.
That’s why compliance needs to be built into agent architecture right away. Mature governance frameworks can scale with an agent ecosystem rather than trying to secure systems designed for speed over control.
Agentic systems consume compute resources continuously through maintaining context, processing multi-turn conversations, and executing reasoning chains that can span thousands of tokens per single decision.
The math is brutal. A customer support agent that looks efficient at 100 tokens per interaction can easily use 2,000–5,000 tokens when the scenario requires multiple tool calls, context retrieval, and multi-step reasoning. Multiply that by enterprise-scale volumes and you can rack up monthly token bills that dwarf even your infrastructure spend.
CPU and GPU utilization follow the same compounding pattern. Every extra thousand tokens is more GPU time. At scale, those seemingly small token decisions become one of your biggest cost line items. Even an “idle” agent can still consume resources through polling, background workflows, state management, monitoring, and context upkeep.
This is exactly why infrastructure and tooling are levers, not afterthoughts. You control token burn by controlling orchestration design, context strategy, caching, routing, evaluation discipline, and the guardrails that prevent looping and runaway workflows.
Architectural intelligence is the focus of cost optimization in agentic AI. The choices you make here either compound efficiency, or compound regret.
Tough truth time: Using the newest, shiniest, most advanced possible engine for every task isn’t the way to go.
Most agent decisions don’t need heavyweight reasoning. Configure your agents to use lightweight models for routine decisions and keep expensive, large language models (LLMs) for more complex scenarios that truly need advanced reasoning.
Fine-tuned, domain-specific engines often outperform larger general-purpose models while consuming fewer tokens and computational resources. This is what happens when architecture is designed intentionally. DataRobot makes this operational by turning model evaluation and routing into an architectural control, not a developer preference — which is the only way this works at enterprise scale.
Infrastructure that scales with demand, not peak capacity, is necessary for controlling agentic AI costs. Auto-scaling and serverless architectures eliminate waste from over-provisioned resources while keeping performance humming during demand spikes.
Kubernetes configurations that understand agentic workload patterns can deliver 40–60% infrastructure savings since agent workloads have predictable patterns (higher during business hours, lower overnight, and spikes during specific business events).
This is where practitioner teams get ruthless: They treat idle capacity as a design bug. DataRobot syftr is built for that reality, helping teams right-size and optimize infrastructure so experimentation and production don’t inherit runaway cloud habits.
Off-peak optimization offers more savings opportunities. Schedule non-urgent agent tasks during low-cost periods, pre-compute common responses, and use spot instances for development and testing workloads. These strategies can reduce infrastructure costs without affecting user experience — as long as you design for them instead of bolting them on.
Open source frameworks like LangChain, AutoGen, and Haystack provide production-ready orchestration capabilities without the licensing costs of commercial alternatives.
Here’s the catch: Open source gives you building blocks, but doesn’t give you enterprise-grade observability, governance, or cost control by default. DataRobot complements these frameworks by giving you the control plane — the visibility, guardrails, and operational discipline required to run agentic AI at scale without duct tape.
Commercial agent platforms can charge $2,000–$50,000+ per month for features that open source frameworks provide for the cost of infrastructure and internal development. For enterprises with technical capability, this can lead to substantial long-term savings.
Open source also provides flexibility that commercial solutions often lack. You can customize orchestration logic, integrate with existing systems, and avoid vendor lock-in that becomes expensive as your agent ecosystem scales.
Manual processes collapse under agentic complexity. Automation saves you time and reduces costs and risks, enabling reliable scaling.
Automated evaluation pipelines test agent performance across multiple scenarios to catch issues before they reach production. CI/CD for prompts and configurations accelerates iteration without increasing risk.
Regression testing becomes vital when agents make autonomous decisions. Automated testing frameworks can simulate thousands of scenarios and validate that behavior remains consistent as you improve the system. This prevents the expensive rollbacks and emergency fixes that come with manual deployment processes — and it keeps “small” changes from becoming million-dollar incidents.
Infrastructure isn’t a supporting actor in agentic AI. It’s a significant chunk of the total cost-savings opportunity, and the fastest way to derail a program if ignored. Getting this right means treating infrastructure as a strategic advantage rather than another cost center.
Caching strategies designed for agentic workloads deliver immediate cost benefits. Agent responses, context retrievals, and reasoning chains often have reusable components. And sometimes, too much context is a bad thing. Intelligent caching can reduce compute costs while improving response times.
This goes hand in hand with pipeline optimization, which focuses on eliminating redundant processing. Instead of running separate inference flows for each agent task, build shared pipelines multiple agents can use.
Your deployment model choice (on-prem, cloud, or hybrid) has massive cost implications.
Here’s your optimization checklist:
The build-versus-buy decision will define both your cost structure and competitive advantage for years. Get it wrong, and you’ll either overspend on unnecessary features or under-invest in capabilities that determine success.
Building your own solution makes sense when you have specific requirements, technical capabilities, and long-term cost optimization goals. Custom development might cost $200,000–$300,000 upfront, but offers complete control and lower operational costs. You own your intellectual property and can optimize for your specific use cases.
Buying a pre-built platform provides faster time-to-market and lower upfront investment. Commercial platforms typically charge $15,000–$150,000+ annually but include support, updates, and proven scalability. The trade-off is vendor lock-in and ongoing licensing costs that grow as you scale.
Hybrid approaches allow enterprises to build core orchestration and governance capabilities while taking advantage of commercial solutions for specialized functions. This balances control with speed-to-market.
| Factor | High | Medium | Low |
| Technical capability | Build | Hybrid | Buy |
| Time pressure | Buy | Hybrid | Build |
| Budget | Build | Hybrid | Buy |
| Customization needs | Build | Hybrid | Buy |
Cost discipline cannot be bolted on later. It’s a signal of readiness and a priority that needs to be embedded into your development lifecycle from day one — and frankly, it’s one of the fastest ways to tell whether an organization is ready for agentic AI or just excited about it.
This is how future-forward enterprises move fast without breaking trust or budgets.
Cost control and performance can coexist. But only if you stop treating cost like a finance problem and start treating it like an engineering requirement.
Your highest-impact optimizations are made up of a few key areas:
But optimization isn’t a one-time effort. Build continuous improvement into operations through regular cost audits, optimization sprints, and performance reviews that balance efficiency with business impact. The organizations that win treat cost optimization as a competitive advantage — not a quarterly clean-up effort when Finance comes asking.
DataRobot’s Agent Workforce Platform addresses these challenges directly, unifying orchestration, observability, governance, and infrastructure control so enterprises can scale agentic AI without scaling chaos. With DataRobot’s syftr, teams can actively optimize infrastructure consumption instead of reacting to runaway spend after the fact.
Learn how DataRobot helps AI leaders deliver outcomes without excuses.
Why is agentic AI more expensive than traditional AI or ML?
Agentic AI is costlier because it does more than return a single prediction. Agents reason through multi-step workflows, maintain context, call multiple tools, and act across systems. That means more model calls, more infrastructure running continuously, and more governance and monitoring to keep everything safe and compliant.
Where do most teams underestimate their agentic AI costs?
Most teams focus on model and GPU pricing and underestimate operational costs. The big surprises usually come from monitoring and debugging overhead, token-heavy conversations and loops, and late-stage governance work that has to be added after agents are already in production.
How do I know if my agentic AI use case is actually worth the cost?
Use a dollar-per-decision view instead of raw infrastructure numbers. For each decision, compare total cost per decision against the value created, such as labor saved, faster resolution times, or revenue protected. If the value per decision does not clearly exceed the cost, you either need to rework the use case or simplify the agent.
What are the fastest ways to cut costs without hurting performance?
Start by routing work to lighter or fine-tuned models for routine tasks, and reserve large general models for complex reasoning. Then, tighten your infrastructure with auto-scaling, caching, and better job scheduling, and turn on automated evaluation so you catch regressions before they trigger expensive rollbacks or support work.How can a platform like DataRobot help with cost control?
A platform like DataRobot helps by bringing observability, governance, and infra controls into one place. You can see how agents behave, what they cost at a decision level, and where they drift, then adjust models, workflows, or infra settings without stitching together multiple tools. That makes it easier to keep both spend and risk under control as you scale.
The post Balancing cost and performance: Agentic AI development appeared first on DataRobot.
As great as your AI agents may be in your POC environment, that same success may not make its way to production. Often, those perfect demo experiences don’t translate to the same level of reliability in production, if at all.
Taking your agents from POC to production requires overcoming these five fundamental challenges:
Building a reliable agent starts by converting vague business goals, such as “improve customer service,” into concrete, quantitative evaluation thresholds. The business context determines what you should evaluate and how you will monitor it.
For example, a financial compliance agent typically requires 99.9% functional accuracy and strict governance adherence, even if that comes at the expense of speed. In contrast, a customer support agent may prioritize low latency and economic efficiency, accepting a “good enough” 90% resolution rate to balance performance with cost.
To reach production readiness, you need to evaluate multiple agentic workflows across different combinations of large language models (LLMs), embedding strategies, and guardrails, while still meeting strict quality, latency, and cost objectives.
Evaluation extends beyond functional accuracy to cover corner cases, red-teaming for toxic prompts and responses, and defenses against threats such as prompt injection attacks.
This effort combines LLM-based evaluations with human review, using both synthetic data and real-world use cases. In parallel, you assess operational performance, including latency, throughput at hundreds or thousands of requests per second, and the ability to scale up or down with demand.
Tracing the execution of agent workflows step by step allows you to understand why an agent behaves the way it does. By making each decision, tool call, and handoff visible, you can identify root causes of unexpected behavior, debug failures quickly, and iterate toward the desired agentic workflow before deployment.
Monitoring deployed agents in production with real-time alerting, moderation, and the ability to intervene when behavior deviates from expectations is crucial. Signals from monitoring, along with periodic reviews, should trigger re-evaluation so you can iterate on or restructure agentic workflows as agents drift from desired behavior over time. And trace root causes of these easily.
You need to apply governance controls at every stage of agent development and deployment to manage operational, security, and compliance risks. Treating governance as a built-in requirement, rather than a bolt-on at the end, ensures agents remain safe, auditable, and compliant as they evolve.
Letting success hinge on hope and good intentions isn’t good enough. Strategizing around this framework is what separates successful enterprise artificial intelligence initiatives from those that get stuck as a proof of concept.
As Agentic AI moves beyond POCs to production systems to automate enterprise workflows, their execution and outcomes will directly impact business operations. The waterfall effects of agent failures can significantly impact business processes, and it can all happen very fast, preventing the ability of humans to intervene.
For a comprehensive overview of the principles and best practices that underpin these enterprise-grade requirements, see The Enterprise Guide to Agentic AI
Before rolling out agents, organizations need confidence in reliability across multiple dimensions, each addressing a different class of production risk.
Reliability at the functional level depends on whether an agent correctly understands and carries out the task it was assigned. This involves measuring accuracy, assessing task adherence, and detecting failure modes such as hallucinations or incomplete responses.
Operational reliability depends on whether the underlying infrastructure can consistently support agent execution at scale. This includes validating scalability, high availability, and disaster recovery to prevent outages and disruptions.
Operational reliability also depends on the robustness of integrations with existing enterprise systems, CI/CD pipelines, and approval workflows for deployments and updates. In addition, teams must assess runtime performance characteristics such as latency (for example, time to first token), throughput, and resource utilization across CPU and GPU infrastructure.
Secure operation requires that agentic systems meet enterprise security standards. This includes validating authentication and authorization, enforcing role-based access controls aligned with organizational policies, and limiting agent access to tools and data based on least-privilege principles. Security validation also includes testing guardrails against threats such as prompt injection and unauthorized data access.
Effective governance requires a single source of truth for all agentic systems and their associated tools, supported by clear lineage and versioning of agents and components.
Compliance readiness further requires real-time monitoring, moderation, and intervention to address risks such as toxic or inappropriate content and PII leakage. In addition, agentic systems must be tested against applicable industry and government regulations, with audit-ready documentation readily available to demonstrate ongoing compliance.
Sustainable deployment depends on the economic viability of agentic systems. This includes measuring execution costs such as token consumption and compute usage, assessing architectural trade-offs like dedicated versus on-demand models, and understanding overall time to production and return on investment.
Pre-deployment evaluation alone is not sufficient to ensure reliable agent behavior. Once agents operate in production, continuous monitoring becomes essential to detect drift from expected or desired behavior over time.
Monitoring typically focuses on a subset of metrics drawn from each evaluation dimension. Teams configure alerts on predefined thresholds to surface early signals of degradation, anomalous behavior, or emerging risk. Monitoring provides visibility into what is happening during execution, but it does not on its own explain why an agent produced a particular outcome.
To uncover root causes, monitoring must be paired with execution tracing. Execution tracing exposes:
This visibility extends to relevant metrics such as accuracy or latency at both the input and output of each step, enabling effective debugging, faster iteration, and more confident refinement of agentic workflows.
And finally, governance is necessary at every phase of the agent lifecycle, from building and experimentation to deployment in production.
Governance can be classified broadly into 3 categories:
Together, monitoring, tracing, and governance form a continuous control loop for operating agentic systems reliably in production.
Monitoring and tracing provide the visibility needed to detect and diagnose issues, while governance ensures ongoing alignment with security, operational, and regulatory requirements. We will examine governance in more detail later in this article.
Many of the evaluation and monitoring practices used today were designed for traditional machine learning systems, where behavior is largely deterministic and execution paths are well defined. Agentic systems break these assumptions by introducing autonomy, state, and multi-step decision-making. As a result, evaluating and operating agentic tools requires fundamentally different approaches than those used for classic ML models.
Classic ML system evaluation is rooted in determinism and bounded behavior, as the system’s inputs, transformations, and outputs are largely predefined. Metrics such as accuracy, precision/recall, latency, and error rates assume a fixed execution path: the same input reliably produces the same output. Observability focuses on known failure modes, such as data drift, model performance decay, and infrastructure health, and evaluation is typically performed against static test sets or clearly defined SLAs.
By contrast, agentic tool evaluation must account for autonomy and decision-making under uncertainty. An agent does not simply produce an output; it decides what to do next: which tool to call, in what order, and with what parameters.
As a result, evaluation shifts from single-output correctness to trajectory-level correctness, measuring whether the agent selected appropriate tools, followed intended reasoning steps, and adhered to constraints while pursuing a goal.
Agentic systems by design are complex multi-component systems, consisting of a combination of large language models and other tools, which may include predictive AI models. They achieve their outcomes using a sequence of interactions with these tools, and through autonomous decision-making by the LLMs based on tool responses. Across these steps and interactions, agents maintain state and make decisions from accumulated context.
These factors make agentic evaluation significantly more complex than that of predictive AI systems. Predictive AI systems are evaluated simply based on the quality of their predictions, whether the predictions were accurate or not, and there is no preservation of state. Agentic AI systems, on the other hand, need to be judged on quality of reasoning, consistency of decision-making, and adherence to the assigned task. Additionally, there is always a risk of errors compounding across multiple interactions due to state preservation.
Agentic evaluation also places far greater emphasis on governance, safety, and cost. Because agents can take actions, access sensitive data, and operate continuously, evaluation must track lineage, versioning, access control, and policy compliance across entire workflows.
Economic metrics, such as token usage, tool invocation cost, and compute consumption, become first-class signals, since inefficient reasoning paths translate directly into higher operational cost.
Agentic systems preserve state across interactions and use it as context in future interactions. For example, to be effective, a customer support agent needs access to previous conversations, account history, and ongoing issues. Losing context means starting over and degrading the user experience.
In short, while traditional evaluation asks, “Was the answer correct?”, agentic tool evaluation asks, “Did the system act correctly, safely, efficiently, and in alignment with its mandate while reaching the answer?”
As enterprises adopt complex, multi-agent autonomous AI workflows, effective evaluation requires more than just accuracy. Metrics and frameworks must span functional behavior, operational efficiency, security, and economic cost.
Below, we define four key categories for agentic workflow evaluation necessary to establish visibility and control.
Functional metrics measure whether the agentic workflow performs the task it was designed for and adheres to its expected behavior.
Core functional metrics:
Why these matter: Functional metrics validate whether agentic workflows solve the problem they were built to solve and are often the first line of evaluation in playgrounds or test environments.
Operational metrics quantify system efficiency, responsiveness, and the use of computational resources during execution.
Key operational metrics
Why these matter: Operational metrics ensure that workflows not only work but do so efficiently and predictably, which is critical for SLA compliance and production readiness.
Security metrics evaluate risks related to data exposure, prompt injection, PII leakage, hallucinations, scope violation, and control access within agentic environments.
Security controls & metrics
Why these matter: Agents often process sensitive data and can interact with enterprise systems; security metrics are essential to prevent data leaks, abuse, or exploitation.
Economic metrics quantify the cost efficiency of workflows and help teams monitor, optimize, and budget agentic AI applications.
Common economic metrics
Why these matter: Economic metrics are crucial for sustainable scale, cost governance, and showing business value beyond engineering KPIs.
Governance and compliance measures ensure workflows are traceable, auditable, compliant with regulations, and governed by policy. Governance can be classified broadly into 3 categories.
Governance in the face of:
Fundamentally, they have to be ingrained in the entire agent development and deployment process, as opposed to being bolted on afterwards.
Ensuring security policy enforcement requires tracking and adhering to organizational policies across agentic systems.
Tasks include, but are not limited to, validation and enforcement of access management through authentication and authorization that mirror broader organizational access permissions for all tools and enterprise systems that agents access.
It also includes setting up and enforcing robust, auditable approval workflows to prevent unauthorized or unintended deployments and updates to agentic systems within the enterprise.
Ensuring operational risk governance requires tracking, evaluating, and enforcing adherence to organizational policies such as privacy requirements, prohibited outputs, fairness constraints, and red-flagging instances where policies are violated.
Beyond alerting, operational risk governance systems for agents should provide effective real-time moderation and intervention capabilities to address undesired inputs or outputs.
Finally, a critical component of operational risk governance involves lineage and versioning, including tracking versions of agents, tools, prompts, and datasets used in agentic workflows to create an auditable record of how decisions were made and to prevent behavioral drift across deployments.
Ensuring regulatory risk governance requires validating that all agentic systems comply with applicable industry-specific and government regulations, policies, and standards.
This includes, but is not limited to, testing for compliance with frameworks such as the EU AI Act, NIST RMF, and other country- or state-level guidelines to identify risks including bias, hallucinations, toxicity, prompt injection, and PII leakage.
Governance metrics reduce legal and reputational exposure while meeting growing regulatory and stakeholder expectations around trustworthiness and fairness. They provide enterprises with the confidence that agentic systems operate within defined security, operational, and regulatory boundaries, even as workflows evolve over time.
By making policy enforcement, access controls, lineage, and compliance continuously measurable, governance metrics enable organizations to scale agentic AI responsibly, maintain auditability, and respond quickly to emerging risks without slowing innovation.
Agentic AI introduces a fundamentally new operating model for enterprise automation, one where systems reason, plan, and act autonomously at machine speed.
This enhanced power comes with risk. Organizations that succeed with agentic AI are not the ones with the most impressive demos, but the ones that rigorously evaluate behavior, monitor systems continuously in production, and embed governance across the entire agent lifecycle. Reliability, safety, and scale are not accidental outcomes. They are engineered through disciplined metrics, observability, and control.
If you’re working to move agentic AI from proof of concept into production, adopting a full-lifecycle approach can help reduce risk and improve reliability. Platforms such as DataRobot support this by bringing together evaluation, monitoring, tracing, and governance to give teams better visibility and control over agentic workflows.
To see how these capabilities can be applied in practice, you can explore a free DataRobot demo.
The post Production-ready agentic AI: key challenges and solutions appeared first on DataRobot.
A core objective of RoboCup is to promote and advance robotics and AI research through the challenges offered by its various leagues. The ultimate goal of the soccer competition is that, by 2050, a team of fully autonomous humanoid robots will defeat the most recent winner of the FIFA World Cup. To bring this vision closer to reality, the RoboCup Federation has announced several changes to the leagues. We spoke with Alessandra Rossi, a trustee who has been involved in the humanoid soccer league for many years, to learn more.
I am Alessandra Rossi from the University of Naples “Federico II”, where I am an Assistant Professor of Computer Science. I began working with and collaborating in RoboCup in 2016, when I started my PhD at the University of Hertfordshire in the UK. I am still affiliated with the University of Hertfordshire, as I remain a member of the humanoid KidSize team Bold Hearts, the longest continuously active team in the UK. After a few years, I became the team leader of Bold Hearts.
In 2019, I became a member of both the Technical Committee and the Organizing Committee of the Humanoid League. After serving on the Technical Committee for two years, I was elected to the Executive Committee of the Humanoid League. In 2025, I was elected to the Board of Trustees for the first time.
Over the years, I have steadily increased my involvement and commitment to RoboCup. I have always sought to actively engage the RoboCup community, both during competitions and outside of competition periods. I also work to encourage engagement between the major and junior leagues and to participate in regional RoboCup events.
While working at the University of Hertfordshire as a Visiting Lecturer, we launched an online module that uses RoboCup as a benchmark for teaching robotics to undergraduate students. The module is still running. I initially served as the module leader, and this role has since been taken over by our Bold Hearts teammate, Bente Riegler.
Last year, Maike Paetzel-Prüsmann, Merel Keijsers, and I (as lead authors), in collaboration with several trustees and many members from different leagues, published a paper on the current and future challenges in humanoid robotics. The paper was published in Autonomous Robots and is, to the best of my knowledge, the first to involve such a large and diverse group of contributors from across the RoboCup leagues. It discusses research within RoboCup and the collaboration and synergies between the leagues.
Group photo of the humanoid league teams at RoboCup 2025.
The 2050 goal of the RoboCup Federation, as many people are probably aware, is for a team of humanoid robots to play against the winners of the FIFA World Cup. To achieve this, it is necessary to push further in that direction. One of the key changes, therefore, will be a stronger focus on humanoid robots.
Another major change will be the merger of the Standard Platform League (SPL) and the KidSize Humanoid League. This merged league will have the freedom to define its exact format and to develop a new roadmap that aligns the entire league toward a shared objective. While the 2050 goal itself remains unchanged, the path toward achieving it will need to be adjusted.
It is crucial to continue fostering the engagement of teams in the leagues that will be affected by these changes. At the same time, we must recognize that technology is advancing rapidly. Over the past year, in particular, we have seen significant progress in both hardware platforms and large language models. As RoboCup serves as a global benchmark for robotics research, we should continuously strive to advance technology and research—while still having fun.
Soccer is the complex task and behavior we are studying, and it is complex in many dimensions: from physical control and robot motion, to communication and strategy, and even human-like interactions. These include responding to the referee’s whistle, verbal and non-verbal communication among team members, interactions with the coach, and communication with the referee. All of these aspects will ultimately need to be incorporated into the humanoid league.
I believe we will see a mix of different robots. With the three companies currently sponsoring RoboCup, we have already seen that their robots can achieve a wide range of behaviors, and there have been significant improvements in robot control. Some of these robots can walk very quickly—almost to the point of running.
Initially, there may be the possibility of multiple teams using the same platform. However, we must keep in mind that both hardware and software can become obsolete very quickly, so we need to remain open to multiple options. A robot that is state of the art today may no longer be so in a year or two. As a result, committing to a single standard platform could limit future progress.
For this reason, the current idea is to remain open to multiple platforms. Many teams already have excellent custom-built robots, and further improvements to these platforms should be encouraged. That said, the exact structure has not yet been decided, and these decisions will be made in consultation with the teams. It is important to give the RoboCup community the time it needs to adapt and move forward.
One major change is that we have added extra robots to each team. Previously, teams played with just two robots per side, but matches are now played three versus three.
Another important improvement is the reduced presence of humans on the field. There is no longer a handler assigned to each robot. In the past, a team member had to walk behind the robots in case they fell and risked being damaged.
I have actually played in a match against the winning humanoid team. Naturally, the human team won, but it was an enjoyable and very interesting game, as the robots were surprisingly fast.
Yes, it depends on which league you are part of. The reactions have been a mix of excitement and passion. Of course, everyone is keen to see improvements, and participants have always been prepared for changes to the rules and the league structures. However, there are still some open questions, and teams are waiting to see how things will evolve. Tomorrow, there will be a meeting with the President and several trustees to address questions raised by the leagues.
The overall direction of RoboCup, guided by the 2050 goal has not changed. Each league has been extremely valuable and has contributed in different ways toward achieving that goal. RoboCup has also been immensely valuable for robotics research more broadly. Beyond being fun, the challenges involved in making robots play soccer are extraordinarily complex. The research and solutions developed within RoboCup can be applied to many other fields and applications.
|
Alessandra is Assistant Professor at the University of Naples Federico II, Italy. Her PhD thesis was part of the Marie Sklodowska-Curie Research ETN SECURE project at the University of Hertfordshire (UK). Her research interests include Human–(Multi) Robot Interaction, social robotics, trust, XAI, multi-agent systems and user profiling. She is Project Manager and co-supervisor of the MSCA PERSEO (955778), TRAIL (101072488) and SWEET (101168792). She is also co-PI of the project ERROR (FA8655-23-1-7060), and part of several national and international projects. Alessandra is also trustee member of RoboCup Federation, and member of the Humanoid League team called Bold Hearts. She is Chair of the IEEE P3108 |
“Study Design”, and a member of the “Appendix” groups, she is Program Chair of IEEE RO-MAN 2027, she has been Robotic Challenge Chair at ICSR 2025, Special Session Chair of IEEE RO-MAN 2024, Publicity Chair of IEEE RO-MAN 2022 and 2023, Organising Chair of the 26th RoboCup International Symposium 2023, and she is on the program committee of several international conferences on human–robot interaction and artificial intelligence.