AgiBot and Tsinghua train 96 robots on their own runs; packaging success hits 98.3%

Tsinghua University, AgiBot and Infinigence AI presented Collective Experience Learning (CEL) on Oct 30, a method that improves a shared robot policy from the fleet's own executions instead of fresh expert demonstrations, according to their arXiv preprint (2610.25736) and AgiBot's announcement as reported by 36Kr.

Operators only label each episode as a success or failure and reset the scene; a failure-guided retrieval step called FERT reuses recorded actions from successful runs, and a fault-tolerant asynchronous system called FATS links robots, cloud inference and training.

On five real manipulation tasks the policies reached success rates of 93.3% to 99.2%, measured over 120 trials per task on robots that did not collect training data. In multi-stage phone packaging, growing the training fleet from 16 to 96 robots raised success from 86.7% to 98.3%, the paper says.

36Kr also reported that CEL beat the average of six human teleoperators by up to 16.7 points, a comparison the abstract does not state. MEKAWIRE context: AgiBot said on Sept 24 that it had produced its 20,000th robot, the size of installed base this kind of fleet learning depends on.