Embodied Intelligence Collective Experience Learning (CEL) Framework Released: Robots Teach Robots to Perform Tasks

Cover News reporter Ouyang Hongyu

Nearly 100 robots, through mutual supervised learning, eventually all became “skilled workers.” This is not a scene from a science fiction movie, but has already become reality.

On October 30, the Collective Experience Learning (CEL) paradigm for robot collective experience learning was released in Shanghai. According to reports, robots trained based on this paradigm have achieved policy success rates exceeding the average level of human remote operators in five real tasks: flower arranging, folding clothes, charging phones, assembling stools, and packaging phones, with the highest surpassing by 16.7 percentage points. It is noteworthy that, within the same physical training time, as the number of robots participating in learning increases, the success rate can also increase synchronously. This means that the increase in the number of robots is the first condition that directly translates into an improvement in task execution capability.

It is understood that this paradigm was jointly developed by AGIBOT, Tsinghua University, and Wuwen Xinqung, aiming to reconstruct the source of robot experience acquisition, namely, learning materials no longer only come from humans but from the real work trajectories of robot peers. In this paradigm, the role of humans changes accordingly: from constantly providing actions and assisting exploration to judging results.

Taking phone packaging as an example, according to experimental data, 96 robots continuously packaged phones 120 times, failing only 2 times. It is noteworthy that, within the same physical training time, when the number of robots participating in learning expanded from 16 to 96, the success rate rose from 86.7% to 98.3%.

The emergence of the CEL paradigm stems from the scale effect of robot training. Industry insiders analyzed that, for cutting-edge general models represented by GPT-6 Astra, zero-shot capabilities are already quite impressive, but the competition for deployment is reliability; even the strongest models will repeatedly make mistakes in contact interaction and long-process execution. The industry's solution is human demonstration, but this has a structural bottleneck: this model is suitable for small-scale laboratory verification, but once expanded to a group of hundreds of robots, manual teaching will become the performance ceiling.

From a technical principle perspective, the CEL paradigm solves the problem of “robots teaching robots.” Its research team proposed the Failure-Guided Experience Retrieval Fine-Tuning algorithm FERT, which is the role of distinguishing between success and failure samples; successful trajectories provide learnable actions, while failure scenarios indicate the key directions for learning. For the collective experience of robot training, the research team designed the Fault-Tolerant Asynchronous Training system FATS, treating robots as training resources equal to GPUs, and finally achieved unified scheduling of all training hardware.

It is worth noting that the CEL paradigm also verified a brand-new Scaling path for embodied intelligence on real machines, which is that under the same interaction time, the more machines participating in training, the higher the task execution success rate. The verification of this path also successfully broke through the bottleneck of the HG-Dagger scheme, which relied on human demonstration, to further expand the scale and break through the success rate upper limit.

Speaking of the future of the CEL paradigm, industry insiders analyzed that this collective experience learning paradigm provides a direction for scaling embodied intelligence that is worth further exploration, allowing robot training to jump out of the traditional route that is highly dependent on human demonstration, thereby autonomously accumulating and mutually learning in real operations. In the future, the successful experiences and failure faults encountered by robots in real working conditions can all be deposited as public assets for the entire fleet. A pit stepped into by one robot can be a lesson learned by the entire fleet; effective operations explored by a single unit can be reused by all robots. When every robot can contribute experience, the quantity itself becomes a new dimension of capability growth.