Tsinghua, Infinigence AI, and AgiBot propose the CEL collective experience learning paradigm, with a 97.5% success rate verified on a 96-robot cluster

Recently, Tsinghua University, Infinigence AI, and AgiBot jointly proposed a new paradigm called "Collective Experience Learning" (CEL), and completed large-scale validation in a cluster of 96 real robots.

The core change in this paradigm lies in the object of learning: robots no longer mainly rely on humans demonstrating actions over and over, but instead reuse the experience accumulated by peers in practice, so that the pitfalls one robot has stepped into become paths the entire collective avoids.

This shift targets a specific obstacle in the deployment of embodied intelligence. Robots can be deployed in batches of hundreds or thousands, but experienced operators cannot be "scaled up" accordingly. At present, training robots mainly relies on humans teaching: humans demonstrate actions and robots imitate; when robots make mistakes, humans then correct them manually. As deployment scale rises, the supply of humans becomes the actual constraint on the speed of expansion.

CEL's solution is to change the object of learning from humans to peers. The successful experience a robot explores in real tasks is retained and enters an experience pool shared by the collective, and when other robots encounter similar scenarios, they can retrieve and reuse it. The more robots participate in training, the richer the experience pool becomes, and the faster capabilities improve. The original bottleneck of "not enough humans" is thus transformed into a scale effect where "the more robots, the stronger."

Validation was completed across a set of real tasks, with scenarios covering both home services and industrial production, specifically including flower arranging, clothes folding, phone charging, stool assembly, and phone packaging; the hardware was the AgiBot G2 dual-arm robot, and the base model was π0.

According to data released by the team, two robots that did not participate in training completed 600 tests in total, with 15 failures, for an overall success rate of 97.5%; the success rates on all five tasks exceeded the average level of 6 human operators.

Supporting this result is a dual design at the system level and the algorithm level.

At the system level, the team built a fault-tolerant asynchronous training system, FATS, 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 as soon as it completes a task, without waiting for other robots, so the training process can run continuously; network fluctuations, device disconnections, and hardware failures are also incorporated into the design. In a system test at the scale of 96 robots, the cluster collected 9,834 rounds of execution experience in total, during which 22 robots experienced 30 failures in total, and the machine-time loss caused by failures accounted for only 1.3% of total operating machine time.

At the algorithm level, the team proposed a failure-guided experience-retrieval fine-tuning algorithm, FERT. A single operation may last several minutes and involve thousands of action steps, but the final feedback is often only a single "success" or "failure," which is insufficient to locate which action should be changed. FERT uses two types of experience simultaneously, learning "what should be done" from successful operations and locating "what is wrong" from failed operations. When a robot encounters difficulties, the algorithm traces back to historical successful cases in similar scenarios, concentrating improvements on the links prone to error.

From an industry perspective, the value of this work lies not entirely in the success-rate numbers, but in its attempt to change the cost structure of capability production. Under the paradigm that relies on manual teaching, capability improvement requires a proportional input of manpower, and the larger the scale, the more teaching personnel need to be on site. Collective learning turns experience into an asset that can be reused within the cluster, so the marginal learning cost per robot has a chance to decline.

The division of labor among the three parties is each well-founded: Tsinghua University provides academic methods and algorithm design, Infinigence AI provides AI infrastructure and training systems, and AgiBot provides complete machines and real scenarios. What is validated is not just a set of algorithms, but the feasibility of the chain in which "real scenarios generate experience, collective learning accumulates capabilities, and large-scale deployment feeds back into more scenarios."

When robots are no longer merely execution units each doing their own work, but can become a collective that learns from one another and progresses together, large-scale deployment is not only the endpoint of capability application, but may also become the starting point of continuous capability accumulation. Risk Disclosure and Disclaimer

Markets carry risks, and investment requires caution. This article does not constitute personal investment advice, nor does it take into account the special investment objectives, financial conditions, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investment based on this is at the user's own risk.