Learning from Fleet Experience: A Collective Learning Paradigm for Robots

Learning from Fleet Experience. A Collective Learning Paradigm for Robots. By Chao Yu and forty-one others, from Tsinghua University, AgiBot, and one other institution. Published on the nineteenth of October. No code repository is linked to it.

A deployed robot generates experience every time it executes a task, and in principle it could become more reliable by learning from its own executions. The usual route runs through human corrections. A teleoperator watches the robot and steps in when the policy fails. That demand becomes hard to meet as fleets grow, because the number of robots to watch rises and the people watching do not.

Collective experience learning, or CEL, takes the humans out of that loop. Robots in a deployed fleet share their execution experience, and that shared experience improves a common policy. One robot's failure can then be answered by another robot's success, rather than by someone at a teleoperation console. The fleet stops being a supervision burden that grows with every robot added, and becomes a pool of experience that grows the same way. Each machine contributes what it did, and every machine benefits from the policy that results.

CEL is realised through failure-guided experience retrieval fine-tuning, called FERT. When a robot fails, the context of that failure is used to retrieve relevant trajectory segments from the fleet's shared pool of successful experience. The actions in those retrieved segments then serve as supervision for updating the policy. No additional expert demonstrations are collected, and no corrective actions either. The failure, in other words, is not something to be fixed by hand. It is a query, and the fleet's past successes are where the answer is looked up.

Running this across many machines needs infrastructure, and that is a fault-tolerant asynchronous training system, FATS. It coordinates experience collection, policy optimisation and the distribution of updated policies back to the fleet, and it keeps doing so despite local interruptions within the fleet.

Across five real-world manipulation tasks, the resulting policies reach success rates between ninety-three point three and ninety-nine point two percent. Each task is evaluated over one hundred and twenty trials, on robots that were excluded from collecting the training data.

Fleet size is tested directly. With collection time per robot held fixed, a larger fleet provides more experience, and FERT generally reaches higher success rates across the sizes evaluated. The clearest case is multi-stage phone packaging, where learning with ninety-six robots raises success from seventy-nine point two percent to ninety-eight point three percent.

The claim is about what scales. Human supervision grows with the fleet and becomes the limit. Shared experience grows with the fleet too, but there it is the resource, not the cost. The authors present collective experience as a practical basis for scalable robot learning, and as a way to make policies more reliable during deployment without asking for new expert corrections.