Why Robots Fail in the Real World: Cambridge Professor Advocates Team-Based Intelligence

TMTPOST -- Despite remarkable advances in artificial intelligence (AI) models, real-world robotics continues to lag behind expectations. Robots frequently stumble in collective tasks, reacting too slowly to real-time demands or failing entirely when confronted with unforeseen scenarios.

This issue, known in the field as 「collective intelligence failure,」 has become a major roadblock for robotics researchers and industry practitioners alike.

In a recent opinion piece published in Science Robotics, Amanda Prorok, Professor of Collective Intelligence and Robotics at the University of Cambridge』s Department of Computer Science and Technology, explains why current robotic systems often fail in collaborative environments and calls for a fundamental rethink of how robotic intelligence is designed. Read the full article here.

The Limits of the Single-Model Approach

Most advanced robots today rely on massive, centralized models designed to handle all tasks—navigation, perception, interaction—through a single architecture. According to Professor Prorok, this approach is inherently flawed. 「The classic pursuit of autonomy—where each robot is expected to act independently—is unsuitable for complex, real-world environments,」 she writes.

The reasoning is straightforward: robots rarely operate in isolation. In reality, they must constantly interact with other agents, whether human or machine, to accomplish complex objectives. Current AI models often ignore these interactions, treating collective behavior as incidental rather than essential. Traditional frameworks for robotic autonomy still define intelligence as an isolated, independent property, a perspective that fails to account for the social and collaborative dynamics critical in real-world settings.

Scaling laws in AI exacerbate the problem. As tasks become more complex, the model size and required data grow exponentially. Large monolithic models, with parameters in the millions or billions, demand massive computational resources and energy. Running these models in real time is often infeasible: they require hundreds of gigabytes of memory and suffer from latency issues, making them unsuitable for high-frequency control and responsive robotics. Even on advanced development boards, only the smallest models can approach real-time performance.

Collective Intelligence: Moving Beyond 「One Brain」

Prorok argues that the solution is not to build a single superintelligent robot but to create collectives of specialized agents that collaborate effectively. In other words, intelligence should be distributed across a team rather than centralized in a single machine. Each robot should focus on a specific skill, while collaboration allows the system as a whole to achieve complex behaviors that no single agent could manage.

This approach relies on modular and compositional design for both hardware and software. Robots in a collective can learn from one another, share experiences, and dynamically reorganize at runtime to adapt to task requirements. The result is 「superlinear」 improvement: combined skills of a team outperform the sum of individual abilities.

Social learning within these collectives also enables robots to develop a deeper understanding of their capabilities and limitations. Skills like theory of mind and metacognition—essential for interacting with humans or other robots—cannot be fully acquired by isolated agents. Instead, they emerge through collaboration, where robots learn when to act independently and when to coordinate.

Experience sharing also reduces risk. In robotics, collecting physical data is costly and potentially dangerous. By distributing knowledge across a collective, robots can avoid repeating hazardous actions, mitigate catastrophic forgetting, and accelerate the overall learning process.

The Challenges of Building Robot Collectives

While the concept of robot collectives is promising, several key hurdles remain:

1. How to Collaborate: Effective robot communication is a significant technical challenge. Most robot-to-robot networks rely on narrowband communication, making it difficult to determine 「what to communicate, when, and with whom.」 Some researchers have experimented with differentiable communication channels or graph neural networks to plan collaboration, but these methods are still in early stages.

2. How to Implement: Designing robots capable of handling diverse and sometimes non-overlapping tasks is difficult. Concepts such as the 「hybrid robot」 paradigm remain underdeveloped. Researchers are exploring solutions inspired by ensemble models, mixture of experts, hypernetworks, and hierarchical learning, but real-time integration of specialized skills is still an open problem.

3. How to Evaluate: Performance metrics are often simplistic, focusing on learning loss or the success of individual robots rather than team-level outcomes. Current evaluation frameworks rarely account for collective resilience, adaptability, or performance in dynamic, multi-agent environments. Without robust standards, robots may excel in isolated tests but fail when teamwork is essential.

Professor Prorok emphasizes that while AI technologies are advancing rapidly, breakthroughs in robotics will require addressing these foundational challenges rather than chasing short-term gains. True robotic intelligence will emerge not from singular, monolithic models but from systems where collaboration, specialization, and adaptability are central.

In practical terms, the robots of the future will function more like teams of humans than isolated machines. Each unit will contribute specialized skills while continuously interacting and learning from its peers. Only then can robots be expected to operate reliably in the unpredictable, dynamic conditions of the real world.

This collective intelligence approach represents a paradigm shift for robotics. It moves away from the notion that one super brain can solve all problems and toward a vision of distributed, adaptable, and socially aware robotic systems. For researchers, engineers, and investors in robotics, the message is clear: collaboration, not size alone, is the key to unlocking the next generation of intelligent machines.

Reference: Amanda Prorok, Collective Intelligence in Robotics: Rethinking Autonomy, Science Robotics, 2025.