Article · AI translation
Let Robots Become One Another’s “Teachers”
Recently, Tsinghua University, Infinigence AI, and AgiBot jointly proposed a new paradigm called Collective Experience Learning (CEL) and validated it at scale with a fleet of 96 physical robots. CEL allows robots to learn not only from human demonstrations but also from the successful experiences of other robots in the group: when one robot encounters difficulty performing a task, it can look to its peers’ past experiences for a solution and continually improve its performance. By sharing and accumulating experience, the group can improve its capabilities together through ongoing practice, offering a new technical path toward autonomous robot evolution and deployment at scale.
As embodied AI gradually moves from laboratories into the real world, robots must work in increasingly complex environments. For them to do that work well, the key is to keep accumulating real-world experience. The common approach today relies on people to “teach” them: either a person demonstrates an action again and again, or, when a robot makes a mistake, a person manually corrects it.
But there is a central problem: robots can be deployed by the hundreds or thousands, while there are not enough experienced operators to increase their numbers along with the robots. This has become a major bottleneck in the development of embodied AI.
CEL’s approach is to give robots a different source of learning: instead of relying mainly on people to teach them, they learn from their “peers.” Successful approaches discovered by one robot through practice can be retained, shared, and reused by the whole group. Robots therefore need not always wait for a person to teach them; they can learn and grow from the experience the group continually accumulates.
As the number of robots grows, so does the group’s store of experience. A unified learning mechanism then continually optimizes their strategies, allowing the robots to become more capable and smarter the more they are used.
Tests in real-world settings further validated CEL’s learning results.
The experimental settings covered both home services and industrial production, with 5 tasks: flower arranging, clothes folding, phone charging, stool assembly, and phone packaging. The results were striking: 2 robots that had not participated in training completed 600 tests with only 15 failures, for an overall success rate of 97.5%. Their success rate on every one of the 5 tasks also exceeded the average of 6 human operators.
Phone packaging is an industrial task that requires several steps to be completed in sequence. After collective experience learning, the robots’ success rate on this task rose from 79.2% to 98.3%, with only 2 failures in 120 consecutive tests. This shows that the learned strategy was effective on a complex task and sufficiently stable.
The research also observed a scaling trend in which “more robots means stronger performance”: when the number of robots participating in training increased from 16 to 96, the success rate rose from 86.7% to 98.3%. In other words, over the same amount of actual training time, a larger group of training robots can accumulate more real-world experience, which can then continually translate into learning gains and make the whole group smarter.
Putting CEL into practice depends on work at both the “underlying system” and “higher-level algorithm” layers.
At the system level, the team built a “Fault-Tolerant Asynchronous Training System (FATS)” based on RLinf, an open-source framework for large-scale reinforcement learning training. It supports continuous, stable, collaborative training for nearly 100 robots. It coordinates physical robots and computing resources: after completing a task, a robot can immediately upload its operating experience without waiting for other robots, and the training system can keep optimizing the model at any time. The system is also prepared for network fluctuations, device disconnections, and hardware failures common in real-world environments: even if some devices go offline, the fleet can continue operating steadily. In a system test with 96 robots, the fleet collected experience from 9834 execution runs. During the test, 22 robots experienced a total of 30 failures, but the robot time lost to failures accounted for only 1.3% of total operating robot-hours. This shows that the system is robust enough to withstand real-world testing at a scale of 100 robots.
At the algorithm level, the team proposed a “Failure-Guided Experience Retrieval Fine-Tuning algorithm (FERT)” to help robots identify how to improve from task successes and failures. There is a practical difficulty here: a robot operation may last several minutes and involve thousands of action steps, but the final feedback is often just “success” or “failure.” That conveys too little information to show which action needs improvement. FERT’s approach uses both kinds of experience: it learns “what to do” from successful operations and identifies “what went wrong” in failed ones. When a robot encounters difficulty, the algorithm searches past successful experiences for effective approaches in similar situations, giving the robot relevant examples to learn from and gradually improving its capabilities.
The key to developing embodied AI as an industry is more than enabling a robot to “do one thing.” The collective learning paradigm jointly introduced by Tsinghua University, Infinigence AI, and AgiBot offers a path from “improving the capabilities of a single robot” to “continual evolution of the entire group.”
And its value extends well beyond the technology itself.
First, it addresses an obstacle that has long held back embodied AI: “not enough people.” In the past, improving robot capabilities depended heavily on experienced operators providing ongoing demonstrations and manual corrections. People are a scarce resource and can never keep pace with the mass deployment of robots. Collective learning changes the situation by allowing robots to learn from one another: the more robots there are, the faster they accumulate experience and improve their capabilities. The “not enough people” bottleneck becomes a scaling advantage in which “more robots means stronger performance.”
What is especially notable is that this collaboration itself exemplifies cooperation among academia, industry, and research. Tsinghua University works at the academic frontier, Infinigence AI is a leading AI infrastructure company, and AgiBot has complete robots and real-world settings. Each party brings its strengths, together bridging the most difficult stretch between “a method in a paper” and “a capability on a production line,” taking cutting-edge algorithms out of papers and into the real world.
What this validation demonstrated was more than an algorithm: it demonstrated a complete process. Real-world settings generate experience, collective learning turns that experience into capabilities, and deployment at scale feeds back into more settings. Once this process works, other companies and settings can draw directly on it, making the path for growth across the embodied AI industry much clearer.
Ultimately, when robots are no longer merely individual units carrying out tasks but become a group that learns from one another and improves together, deployment at scale may become not just the endpoint for applying their capabilities but also the starting point for their continual evolution. This is a change in how robots learn and a genuine breakthrough in moving embodied AI from technical validation to industrial use.
This article was provided by Infinigence AI and republished by QbitAI with authorization. The views are those of the original author.