Understand this story
The joint team from Tsinghua University, AgiBot, and Infinigence AI proposes Collective Experience Learning (CEL), enabling deployed robots to share execution experience and jointly improve the same policy; its algorithm is implemented as Failure-guided Experience Retrieval Fine-Tuning (FERT), requiring no additional expert demonstrations or corrective actions. According to a 36Kr·AI report, on five real-world tasks including phone packaging, flower arrangement, and clothes folding, the final policy achieved success rates of 93.3% to 99.2%; for the multi-stage phone packaging task, after learning with 96 robots, the success rate rose from 79.2% to 98.3%, averaging 16.7 percentage points higher than human teleoperation. The team also built a fault-tolerant asynchronous training system, FATS; 96 robots collected 9,834 episodes over nearly 8 hours, and 30 failures caused only 1.3% loss of robot time. Currently this work is at the paper and system release stage, with no subsequent progress disclosed.
AI-generated from public reports
Developments
1 confirmed developments
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 22
36Kr · AI
Tsinghua, AgiBot, and Infinigence AI propose Collective Experience Learning (CEL); 96 robots co-evolve and surpass human success rates
A joint team from Tsinghua University, AgiBot, and Infinigence AI proposes Collective Experience Learning (CEL), which enables deployed robots to share execution experience and jointly improve the same policy. The algorithm is implemented as Failure-guided Experience Retrieval Fine-Tuning (FERT) and requires no additional expert demonstrations or corrective actions. On five real-world tasks, including phone packaging, flower arrangement, and clothes folding, the final policy achieved success rates of 93.3% to 99.2%. On the multi-stage phone packaging task, learning with 96 robots raised the success rate from 79.2% to 98.3%, on average 16.7 percentage points higher than human teleoperation. The team also built a fault-tolerant asynchronous training system, FATS; 96 robots collected 9,834 episodes over nearly 8 hours, and 30 failures caused only 1.3% loss of robot time.