Robots Share Experience: 96-Machine Fleet Boosts Packing Success From 79.2% to 98.3%

Deploying robotic fleets at the scale of dozens of devices requires solving a basic practical problem: every robot needs operator oversight whenever failures occur. Researchers have unveiled Collective Experience Learning (CEL) — a paradigm in which all robots in a fleet learn from shared task experience, eliminating the need for constant human intervention.

CEL Architecture

The system consists of two key components. The first is FERT (failure-guided experience retrieval fine-tuning). When a robot encounters a failure, FERT analyzes the context in which it occurred and searches the fleet's accumulated database of successful trajectories for relevant segments. Actions from these successful examples serve as a signal for updating the robot's policy, without needing to collect additional expert demonstrations or require operator correction.

The second component is FATS (fault-tolerant asynchronous training system), designed specifically for the realities of deployed systems. FATS asynchronously coordinates three processes simultaneously: collecting experience from robots, optimizing the shared policy, and distributing the updated policy back to the devices. This asynchronous architecture lets the system keep running even when individual robots fail or communication is disrupted.

Results in Practice

Researchers tested CEL on five real-world manipulation tasks. The results clearly demonstrate a scaling effect: as the fleet grows while per-robot data collection time stays fixed, success rates continuously improve. The effect is most pronounced on the multi-stage phone-packing task: expanding the fleet from 16 to 96 robots raised success from 86.7% to 98.3%. The policy's baseline success rate before applying CEL was 79.2%.

Across the full set of five tasks, CEL achieves success rates in the 93.3–99.2% range. All evaluations were conducted on robots excluded from the training data collection, with 120 trials per task.

The Economics of Scaling

CEL solves a critical deployment problem. The traditional approach requires one operator per new robot to monitor and correct failures. With CEL and FATS, failures instead become training signals for the shared policy, which improves through the fleet's aggregate experience. Every additional robot contributes to improving the policy for all the others. The cost of collecting experience per device stays constant, but a larger fleet simply accumulates more diverse experience for FERT to draw useful examples from. In the phone-packing case, this produces a jump from 79.2% success at initialization to near-complete reliability at maximum scale.