Tsinghua University, together with Infinigence AI and AgiBot, proposes Collective Experience Learning (CEL); 96 robots validate a 97.5% success rate

From learning from humans to learning from robot peers: Tsinghua University, Infinigence AI, and AgiBot propose a new path for the large-scale deployment of embodied intelligence

Friends of QbitAI 2026-10-30 11:48:04

Source: QbitAI

Let robots become each other's "teachers"

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. CEL allows robots to not only rely on human demonstrations for learning, but also to draw on the successful experience of other robots in the collective: when a robot encounters difficulties while executing a task, it can look for solutions in the past experience of its peers and continuously improve its own performance. By sharing and accumulating experience, a robot collective can jointly improve its capabilities through continuous practice, providing a new technical path for promoting the autonomous evolution and large-scale application of robots.

1. Letting robots become each other's "teachers": a new direction for the large-scale development of robots

As embodied intelligence gradually moves from the laboratory to the real world, robots need to work in increasingly complex environments. To make them perform well, the key is to enable them to continuously accumulate real-world experience. The current common approach is to rely on humans to "teach": either a person repeatedly demonstrates actions for the robot to watch, or when the robot makes a mistake, a person manually corrects it.

But there is a core problem here: robots can be deployed in batches of hundreds or thousands, but there are not that many experienced operators, and it is impossible for them to "scale up" alongside the robots. This has become a major bottleneck in the development of embodied intelligence.

CEL's approach is to give robots a different learning target — not mainly relying on human teaching, but on learning from "peers." The successful experience that one robot figures out through practice can be retained and shared and reused by the entire robot collective. In this way, robots do not have to keep waiting for humans to teach them; instead, they can learn and grow on their own from the experience continuously accumulated by the collective.

As the number of robots increases, the experience accumulated by the collective becomes richer, and through a unified learning mechanism that continuously optimizes strategies, the robots' capabilities can become stronger with use and smarter with practice. 2. Nearly a hundred robots "learn as a group," with task success rates all exceeding the human average

Tests in real scenarios further validated the learning effect of CEL.

The experimental scenarios included both home services and industrial production, specifically five tasks: flower arranging, clothes folding, phone charging, stool assembly, and phone packaging. The results were quite impressive: two robots that had not participated in training underwent 600 tests and failed only 15 times, for an overall success rate of 97.5%, and the success rates on all five tasks exceeded the average level of 6 human operators.

Among them, phone packaging is an industrial task that requires completing multiple steps in sequence. After collective experience learning, the robot's success rate on this task increased from 79.2% to 98.3%, failing only 2 times in 120 consecutive tests. This shows that the learned strategies are not only effective on complex tasks but also sufficiently stable.

The study also observed a scaling trend in which "the more robots, the stronger": when the number of robots participating in training increased from 16 to 96, the success rate rose from 86.7% to 98.3%. That is, within the same actual training time, the larger the scale of robots participating in training, the more real experience can be accumulated, and this experience can in turn be continuously converted into learning gains, making the entire collective smarter and smarter. 3. Supporting algorithmic innovation with a flexible, highly fault-tolerant system to close the loop of continuous robot learning

CEL's real-world deployment relies on joint efforts at both the "underlying system" and "upper-layer algorithm" levels.

At the system level, the team built a "Fault-tolerant Asynchronous Training System (FATS)" based on the open-source large-scale reinforcement learning training framework RLinf, supporting nearly a hundred robots in continuous and stable collaborative learning and training. It uniformly schedules real robots and computing resources: after a robot completes a task, it can immediately upload its operational experience without waiting for other robots, and the training system can continue optimizing the model at any time. At the same time, the system is prepared in advance for network fluctuations, device disconnections, and hardware failures that are common in real environments — even if some devices "go offline," the entire cluster can still operate stably. In a system test at the scale of 96 robots, the cluster collected a total of 9,834 rounds of execution experience; during this period, 22 robots experienced a total of 30 failures, but the robot time lost due to failures accounted for only 1.3% of total operating machine-hours. This shows that the system is robust enough to withstand the test of a real environment at the scale of a hundred robots.

At the algorithm level, the team proposed the "Failure-guided Experience Retrieval Fine-tuning algorithm (FERT)" to help robots find directions for improvement from task successes and failures. There is a practical difficulty here: a single robot operation may last several minutes and involve thousands of action steps, but the final feedback is often just a single word, "success" or "failure," which carries too little information to identify which action should be improved. FERT's approach is to make use of both types of experience: learn "what should be done" from successful operations, and find "what was done wrong" from failed operations. Once a robot encounters difficulties, the algorithm searches past successful experience for effective practices in similar scenarios, so that the robot can focus on learning from them, and its capabilities improve bit by bit in this way. 4. Industry-academia-research collaboration to explore a new future for the industrialization of embodied intelligence

The industrialization of embodied intelligence hinges not only on making robots "able to do one thing." The collective learning paradigm jointly introduced by Tsinghua University, Infinigence AI, and AgiBot offers a path from "improving the capability of a single robot" to "the continuous evolution of the entire collective."

And its value goes far beyond the technology itself.

It first addresses an obstacle that has long held back the development of embodied intelligence — "not enough people to go around." In the past, improving robot capabilities depended heavily on experienced operators continuously teaching and manually correcting errors. But people are a scarce resource and can never keep pace with the speed of batch robot deployment. Collective learning lets robots learn from one another, and the situation changes completely: the more robots there are, the faster experience accumulates and the faster capabilities grow, turning the "not enough people" bottleneck into a scale advantage where "the more robots, the stronger they become."

Even more remarkable is that this collaboration itself is the best footnote to industry-academia-research cooperation. Tsinghua University stands at the academic frontier, Infinigence AI is a leading AI infrastructure company, and AgiBot possesses complete machines and real-world scenarios — each of the three contributes its strengths, precisely bridging the hardest stretch of the road from "methods in papers" to "capabilities on the production line," so that cutting-edge algorithms no longer remain in papers but run all the way into the real world.

And what has been validated this time is not just a set of algorithms, but a complete chain: real-world scenarios generate experience, collective learning consolidates capabilities, and large-scale deployment feeds back into more scenarios. Once this chain runs smoothly, other enterprises and other scenarios can draw on it directly, and the growth path of the entire embodied intelligence industry will become much clearer as a result.

Ultimately, 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 no longer just the endpoint of capability application, but is also expected to become the starting point of continuous capability evolution. This is not only a shift in how robots learn, but also a true breakthrough for embodied intelligence as it moves from technical validation to industrial implementation.

This article is provided by Infinigence AI; QbitAI is authorized to repost it, and the views belong to the original author.