From Learning Humans to Learning Robot Peers: A New Path for Scalable Embodied Intelligence

Recently, Tsinghua University, Wuwen Xinqiong, and AGIBOT jointly proposed a new paradigm called Collective Experience Learning (CEL), and conducted large-scale validation using a swarm of 96 real robots. CEL enables robots to not only learn from human demonstrations but also draw on the successful experiences of other robots in the group: when a robot encounters difficulties during a task, it can seek solutions from the past experiences of its peers, continuously improving its own performance. By sharing and accumulating experiences, the robot group can collectively enhance its capabilities through continuous practice, providing a new technical pathway for promoting the autonomous evolution and large-scale application of robots.

One, letting robots become each other's “teachers”, robot scaled development has a new direction

As embodied intelligence gradually moves from laboratories into the real world, robots need to work in increasingly complex environments. To make them perform well, the key is to let them continuously accumulate real-world experience. The current common approach relies on humans 'teaching' them: either by having humans repeatedly demonstrate actions to them, or by manually correcting them when they make mistakes.

But there is a core issue here: robots can be deployed in hundreds or even thousands, while experienced operators are not in such abundance, making it impossible for them to scale along with the robots. This has become a major bottleneck in the development of embodied intelligence.

The idea behind CEL is to have robots learn from different sources — not primarily relying on humans to teach them, but rather learning from their “peers.” Success experiences gained by one robot through practice can be accumulated and shared, reused by the entire robot population. In this way, robots don’t have to wait for humans to teach them; instead, they can learn and grow by continuously drawing from the accumulated experiences of the group.

As robots become more numerous, the collective experience they accumulate becomes richer. By employing a unified learning mechanism to continuously optimize strategies, the capabilities of the robots will become stronger with use and smarter the more they work.

II. Nearly 100 robots 'team up to learn', task success rates all exceed the human average

The quiz in real-world scenarios further validated the learning effectiveness of CEL.

The experimental scenarios included both household services and industrial production, specifically including flower arrangement, folding clothes, charging a phone, assembling a stool, and packaging a phone. The results were quite impressive: Two robots that had not participated in any training performed 600 tests, failing only 15 times, achieving an overall success rate of 97.5%, and the success rates for all five tasks exceeded the average performance of six human operators.

Among these, phone packaging is an industrial task that requires completing multiple steps sequentially. After group experience learning, the robot's success rate for this task increased from 79.2% to 98.3%, failing only twice in 120 consecutive tests. This demonstrates that the learned strategies are not only effective but also stable for complex tasks.

The study also observed a scaling trend of “the more machines, 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%. In other words, With the same actual training time, the larger the scale of robots participating in training, the more real experience they can accumulate, and this experience can continuously convert into learning gains, making the entire group increasingly intelligent.

III. Support algorithm innovation with a flexible, high fault-tolerant system, and establish a continuous learning loop for robots

CEL can truly take root, relying on both the underlying system and the upper-level algorithms working together.

At the system level, the team based their work on a open-source framework for large-scale reinforcement learning training RLinf Built the ``Fault-Tolerant Asynchronous Training System (FATS)'', supporting continuous stable collaborative learning and training of nearly 100 robots. It unifies real robots with computational resources: after a robot completes a task, it can immediately upload its operational experience, eliminating the need to wait for other robots. The training system can continuously optimize the model at any time. At the same time, common issues such as network fluctuations, device disconnections, and hardware failures in real environments are all anticipated and prepared for ——The system has also made preparations in advance for common network fluctuations, device disconnections, and hardware failures that are often seen in such scenarios —— Even if some devices go offline, the entire cluster can still operate stably. In the system test with a scale of 96 robots, the cluster accumulated 9,834 rounds of execution experience. During this period, 22 robots experienced a total of 30 failures, but the time loss caused by the failures accounted for only 1.3% of the total operating hours. This indicates that the system is sufficiently robust to withstand the real-world test of a 100-robot scale.

At the algorithm level, the team proposed the ``Failure-Guided Experience Retrieval and Fine-Tuning Algorithm (FERT)'', helping robots find improvement directions from task success and failure. Here's a real challenge: a single robot operation may last several minutes and involve thousands of action steps, but the feedback received at the end is often just one sentence, “success” or “failure,” which contains too little information to determine which specific action needs improvement. FERT's approach is to make use of both types of experience: Learn 'what should be done' from successful operations and find 'where things went wrong' from failed ones. When a robot encounters difficulties, the algorithm will search through past successful experiences to find effective approaches from similar scenarios, allowing the robot to focus on learning these methods, thereby gradually improving its capabilities.

IV. Industry-Academia-Research Collaboration, Exploring a New Future for the Industrialization of Embodied Intelligence

The industrialization of embodied intelligence is not just about getting robots to “do one thing well.” This time, the joint release by Tsinghua University, Wuwen Xinqiong, and AGIBOT of the group learning paradigm provides a path from “enhancing the capabilities of individual robots” to “the continuous evolution of the entire group.”

But its value goes far beyond the technology itself.

It first addresses a long-standing bottleneck in the development of embodied intelligence —— “lack of human resources.” In the past, as robot capabilities improved, they heavily relied on experienced operators for continuous demonstration and manual error correction. However, humans are a scarce resource and can never keep up with the speed of robot mass deployment. With group learning, robots can learn from each other, changing the situation Completely different: The more robots there are, the faster experience accumulates and capabilities grow. The bottleneck of 'not enough people' becomes a scale advantage of 'the more robots, the stronger they become.'

What is more remarkable is that this collaboration itself is the best example of the synergy between academia, industry, and research. Tsinghua University represents the forefront of academic research, while Wuqiong is a leading company in AI infrastructure, and AGIBOT possesses complete systems and real-world scenarios. Each party contributes its strengths, precisely bridging the hardest part of the journey from 'methods in papers' to 'capabilities on production lines,' allowing cutting-edge algorithms to no longer remain confined to academic papers but to run directly into the real world.

But this verification is not just about a single algorithm, but a complete end-to-end pipeline: real-world scenarios generate experience, group learning accumulates capabilities, and large-scale deployment feeds back into more scenarios. Once this pipeline is successfully implemented, other companies and other scenarios can directly draw lessons from it, significantly clarifying the growth path of the embodied intelligence industry.

In the end, when robots are no longer just individual execution units but can become entities that learn from each other and progress collectively, large-scale deployment is no longer just the endpoint of capability application, but also has the potential to become the starting point for continuous capability evolution. This is not just a shift in how robots learn, but a true breakthrough for embodied intelligence moving from technical validation to industrial implementation.