AgiBot, together with Tsinghua University and Infinigence AI, releases the Collective Experience Learning (CEL) paradigm

A robot continuously packaged phones 120 times, failing only twice, raising the task success rate from 79.2% to 98.3%. Behind this report card, there was no human demonstrating each unit one by one; the robots' learning experience came from "copying each other's homework." On October 30, AgiBot, together with Tsinghua University and Infinigence AI, officially released the Collective Experience Learning (CEL) paradigm. This paradigm enables robots to achieve collective autonomous learning by sharing successful and failed experiences from real operations. Traditional embodied intelligence training relies heavily on human demonstrations; robots need to accumulate data through extensive manual teleoperation before model training. This approach is still feasible in small-scale experiments, but when the number of robots expands to hundreds, the collection efficiency of human demonstrations may become a bottleneck constraining performance improvement. CEL attempts to change this model: learning materials no longer come only from humans, but from the real operational trajectories of robot peers. When one robot successfully completes an operation, its experience can be stored in a shared experience pool for other robots to learn from; when another robot encounters a failure, the system can, based on the failure scenario, retrieve successful operations under similar working conditions from the experience pool to help the policy optimize its weak points. Thus, the practice results of a single robot can be transformed into learning resources for the entire cluster, and the human role gradually shifts from action demonstrator to task result adjudicator. The result is that in five real tasks—flower arrangement, clothes folding, phone charging, stool assembly, and phone packaging—the policies trained with CEL all exceeded the average level of 6 human teleoperators, leading by as much as 16.7 percentage points. To truly enable robots to learn from each other, two major challenges still need to be solved: sparse feedback in long-process tasks and instability in parallel operation across multiple devices. To address this, the team developed a failure-guided experience retrieval fine-tuning algorithm and a fault-tolerant asynchronous training system. The former uses failure trajectories to locate policy weaknesses, then retrieves matching successful operations from the collective experience pool for focused training; the latter allows robots to upload data independently after completing tasks without waiting for other devices, while also supporting dynamic recovery after a robot goes offline, charges, or malfunctions, preventing individual device anomalies from affecting the entire cluster. Data shows that in a real training experiment with 96 robots, the cluster completed a cumulative 9,834 rounds of task collection, with 30 robot failure events occurring, and the machine-time loss caused by failures accounting for only 1.3% of total operating time. This means that even if some robots temporarily drop out, other devices can continue working and learning. More noteworthy is that the experiment validated a new path for the scaled development of embodied intelligence. An increase in the number of robots can directly translate into improved task execution capability. In the phone packaging task, as the number of robots participating in training increased from 16 to 96, the test success rate rose from 86.7% to 98.3% within the same physical interaction time. The release of the CEL paradigm means that capability improvement in embodied intelligence is no longer entirely limited by the growth of human demonstration data; the robot cluster itself is also expected to become an important source of continuously accumulating experience and enhancing operational capability.