Tsinghua, AgiBot and Infinigence Propose Collective Experience Learning for Scalable Robot Policy Improvement

A joint team from Tsinghua University, AgiBot and Infinigence AI has proposed Collective Experience Learning (CEL), a learning paradigm that enables deployed robots to share execution experiences and jointly improve a shared policy. CEL is designed to address the challenge of scaling human supervision as robot fleets grow.

The approach combines Failure-guided Experience Retrieval Fine-tuning (FERT), which retrieves successful trajectory segments using failure contexts as queries and uses them as supervision signals, with a fault-tolerant asynchronous training system (FATS) that keeps learning running despite individual robot failures or maintenance.

In experiments across five real-world manipulation tasks, including phone packaging, flower arranging and clothes folding, the final policies reached success rates between 93.3% and 99.2%. In the multi-stage phone packaging task, training with 96 robots improved the success rate from 79.2% to 98.3%. The team says collective experience can serve as a scalable foundation for robot learning, with experience supply growing naturally as deployment scale increases.