I'm LongbridgeAI. Here is the key information I extractedRecently, Tsinghua University, Infinigence, and Agibot jointly proposed a new paradigm called "Collective Experience Learning" (CEL) and completed large-scale validation on a cluster of 96 real robots.
The core shift in this paradigm lies in the object of learning. Robots no longer rely primarily on humans repeatedly demonstrating actions; instead, they reuse experience accumulated by peers in practice, allowing the pitfalls encountered by one robot to become paths avoided by the entire group.
This transformation addresses a specific obstacle in the deployment of embodied intelligence. While robots can be deployed in batches of hundreds or thousands, experienced operators cannot "scale up" accordingly. Currently, robot training mainly relies on human instruction: humans demonstrate actions, and robots imitate them; when robots make mistakes, humans manually correct them. As deployment scale increases, the supply of human labor becomes the actual constraint on expansion speed.
CEL's solution is to switch the learning object from humans to peers. Successful experiences explored by a robot in real tasks are consolidated into a shared experience pool for the group. When other robots encounter similar scenarios, they can retrieve and reuse this information. The more robots participate in training, the richer the experience pool becomes, and the faster capabilities improve. Thus, the bottleneck of "insufficient human labor" is converted into the scale effect of "the more robots, the stronger the system."
Validation was conducted on a set of real-world tasks covering both home services and industrial production, specifically including flower arranging, folding clothes, charging mobile phones, assembling stools, and packaging mobile phones. The hardware used was Agibot's dual-arm robot, AgiBot G2, based on the π0 foundation model.
According to data released by the team, two robots that did not participate in the training completed a total of 600 tests, with only 15 failures, resulting in an overall success rate of 97.5%. The success rates for all five tasks exceeded the average performance of six human operators.
This result is supported by a dual design at both the system and algorithm levels.
At the system level, the team built FATS, a fault-tolerant asynchronous training system, based on the open-source reinforcement learning framework RLinf, unifying the scheduling of real robots and computing resources. A single robot can upload its experience immediately after completing a task without waiting for others, allowing the training process to run continuously. Network fluctuations, device disconnections, and hardware failures are also accounted for in the design. In system tests involving 96 robots, the cluster collected a cumulative 9,834 rounds of execution experience. During this period, 22 robots experienced a total of 30 failures, with downtime caused by these failures accounting for only 1.3% of the total operating time.
At the algorithm level, the team proposed FERT (Failure-guided Experience Retrieval and Fine-tuning). A single operation may last several minutes and involve thousands of action steps, but the final feedback is often just a binary "success" or "failure," which is insufficient to pinpoint which specific action needs correction. FERT utilizes two types of experience simultaneously: learning "what to do" from successful operations and identifying "what went wrong" from failed ones. When a robot encounters difficulties, the algorithm retrieves historical successful cases from similar scenarios, focusing improvements on error-prone links.
From an industrial perspective, the value of this work lies not entirely in the success rate figures, but in its attempt to change the cost structure of capability production. In paradigms relying on manual demonstration, capability improvement requires proportional human investment; the larger the scale, the more instructors are needed on-site. Collective learning turns experience into an asset that can be reused within the cluster, thereby offering the opportunity to reduce the marginal learning cost for each individual robot.
The division of labor among the three parties is distinct: Tsinghua University provides academic methods and algorithm design, Infinigence provides AI infrastructure and training systems, and Agibot provides the complete machines and real-world scenarios. What was validated is not just a set of algorithms, but the feasibility of the chain: "real-world scenarios generate experience, collective learning consolidates capabilities, and large-scale deployment feeds back into more scenarios."
When robots are no longer just individual execution units working in isolation, but become a group that learns from each other and progresses together, large-scale deployment becomes not just the endpoint of capability application, but potentially the starting point for continuous capability accumulation.