AgiBot, together with Tsinghua University and Infinigence AI, releases the CEL collective experience learning paradigm

English text is being prepared.

Zhong Lilei, 1 hour ago — Opening a new path for autonomous growth.

Introduction: As the seventh installment of AI WAVE, AgiBot, together with Tsinghua University and Infinigence AI, releases the Collective Experience Learning (CEL) paradigm, continuously implementing the "put robots truly on the job — a full-stack closed loop from data to intelligence" operational intelligence layout. The paradigm connects the complete chain of experience acquisition, experience utilization, and large-scale online learning; with "robot peers serving as each other's source of experience" as the core breakthrough, it turns every success and failure in real operations into shared learning material for the collective, allowing the expansion of robot cluster scale to translate directly into improved task execution reliability, and opening a new path for autonomous growth that frees embodied intelligence from the scaling bottleneck of "expert demonstrations being difficult to increase in sync."

120 consecutive phone-packaging runs with only 2 failures — this is not the report card of a skilled worker, but that of a robot: policy success rate raised from 79.2% to 98.3%, failures reduced from 25 to 2, a decrease of 92%. More critically, the method: these 96 robots did not wait for any human teacher to demonstrate one by one — their action supervision came from each other.

On October 30, as the seventh installment of AI WAVE, AgiBot, together with Tsinghua University and Infinigence AI, officially released the Collective Experience Learning paradigm CEL (Collective Experience Learning). On five real tasks — flower arrangement, clothes folding, phone charging, stool assembly, and phone packaging — the policies trained with CEL all exceeded the average level of 6 human teleoperators, surpassing it by up to 16.7 percentage points — within the same physical training time, the number of robots participating in learning expanded from 16 to 96, and the success rate rose from 86.7% to 98.3%. For the first time, an increase in the number of robots translated directly into improved task execution capability.

CEL paper concept diagram

Frontier general models represented by GPT-6 Astra already have impressive zero-shot capabilities, but what matters in deployment is reliability — even the strongest models repeatedly make errors in contact-rich interaction and long-horizon execution. The industry's solution is human demonstration, but this has a structural bottleneck: this model suits small-scale laboratory validation, but once extended to a cluster of hundreds of robots, human teleoperation becomes the performance ceiling.

The CEL paradigm reconstructs the source of robot experience acquisition: learning material no longer comes only from humans, but from the real operational trajectories of robot peers.

Suppose one robot encounters difficulty when placing a phone, while another robot has previously succeeded in a similar situation. Then the latter's experience can help the shared policy compensate for the former's shortcomings, so that one robot's practice benefits the entire collective.

In this paradigm, the human role changes accordingly: from continuously providing actions and assisting exploration to judging outcomes.

Differences between CEL and previous learning paradigms

But to take "robots teaching robots" from a concept to an engineerable, operational system, two major practical technical challenges must be overcome:

First, the sparse feedback problem. Long-horizon tasks often involve thousands of action steps, yet the final result is only a 0/1 bit of success or failure, making it impossible to locate which step went wrong;

Second, the cluster stability problem. With hundreds of robots operating in parallel, there will be charging resets, network jitter, and occasional hardware failures; an anomaly in an individual device must not interrupt the entire cluster training process.

To address these two pain points, CES introduced the FERT failure-guided experience retrieval fine-tuning algorithm and the FATS fault-tolerant asynchronous training system. FERT algorithm: failures point the direction, successes provide templates

Real-world industrial operational tasks often have long chains; take phone packaging as an example: a complete execution lasts over 3 minutes and contains more than 4,000 action steps, yet the final feedback for the task is only one bit: success recorded as 1, failure as 0. Such feedback can distinguish whether the task is completed, but cannot directly indicate which stages the policy is deficient in, or which successful experiences are worth focusing on learning. To this end, the team proposed the failure-guided experience retrieval fine-tuning algorithm FERT. It distinguishes the roles of success and failure samples: successful trajectories provide learnable actions, while failure scenarios indicate the focus of learning.

FERT algorithm framework diagram

Failure trajectories are used to locate weaknesses: observations near the failure moment of failed tasks are intercepted, jointly encoded with images and joint states, and used to mark the operating conditions the cluster needs to reinforce;

Successful trajectories are used to output templates: all valid successful trajectories generated by the robots are stored in a collective shared experience pool;

Retrieval-prioritized sampling training: using the features of failure scenarios as query conditions, successful operation segments with matching conditions are retrieved from the experience pool, raising the training sampling priority of such samples and concentrating compute on high-frequency error-prone stages;

Policy iteration loop: the updated policy is distributed to all robots, which continue to produce a new round of success and failure trajectories, cycling repeatedly for continuous optimization.

In one sentence: failure is not merely a negative record; it tells the robots which scenarios to focus on learning next. The value of collective experience comes, on the one hand, from continuous data accumulation, and on the other hand, from whether the appropriate experience can be precisely delivered to where the model most needs to learn. FATS system: 96 real robots trained in parallel, fault time loss only 1.3%

Enable 96 robots to continuously accumulate experience while facing the following real-world problems: some execute quickly while others reset slowly; some have network fluctuations while others need charging; and a few robots may even experience hardware failures.

If a local failure required restarting the entire training run, large-scale collaborative training would be out of the question.

To address this problem, the team designed the fault-tolerant asynchronous training system FATS, which continuously and dynamically organizes 100 robots, 50 edge nodes, and 6 cloud servers, stably supporting large-scale real-world online learning. FATS treats robots as training resources equivalent to GPUs, enabling unified scheduling of all training hardware.

1. SwarmLink: asynchronous data streams, no need for everyone to wait

The data plane SwarmLink is responsible for connecting the environment, inference, and training nodes. After each robot completes a task, it independently uploads its task trajectory; the training module can update the policy in real time upon receiving the data, without waiting for all robots to complete a round of tasks. Training and inference components can be flexibly deployed in the cloud or at the edge, adapting to variable network bandwidth conditions and balancing communication overhead against interaction latency.

SwarmLink design

2. HotSwarm: hot-swapping of nodes, in-place recovery from failures

The control plane HotSwarm manages the dynamic entry and exit of nodes and worker processes. It supports in-place restart and recovery after a robot goes offline, charges, or fails, without resetting the entire cluster. In an experiment with 96 robots, the cluster completed a total of 9,834 rounds of task collection; during this period, 30 robot failure events occurred, and the machine-time loss caused by failures accounted for only 1.3% of the cluster's total operating time. A problem with a local device will not interrupt the learning iterations of the entire swarm.

HotSwarm design Five real tasks tested: cluster learning performance surpasses human teleoperators

This real-robot experiment covered five real tasks: flower arrangement, clothes folding, phone charging, stool assembly, and phone packaging, covering typical scenarios such as deformable object manipulation, precision assembly, tool use, and long-process industrial operations. A total of 600 real-robot tests were completed, with total testing time approaching 30 robot-hours.

The control group set up mainstream approaches such as SFT offline fine-tuning and HG-DAgger human-in-the-loop teaching for horizontal comparison:

Success rates on the 5 test tasks

1. Traditional SFT offline fine-tuning achieved an average task success rate of only 36.3%;

2. The HG-DAgger approach, which relies on extensive human intervention, improved this to 81.0%;

3. Under the CEL-FERT paradigm, all five tasks surpassed the average level of 6 human teleoperators, leading by up to 16.7 percentage points, with an overall task success rate of 97.5%.

CEL truly eliminates reliance on human teaching and achieves more stable and more robust task execution. A new Scaling law validated: the more robots, the better the task performance

The most critical value of the CEL paradigm is that it validated a new Scaling path for embodied intelligence on real robots: under the same interaction time, the more robots participating in training, the higher the task success rate.

Real-robot scaling curve

The team gradually expanded the scale of the robot cluster participating in training. Within the same physical time, the FERT algorithm showed a significant scaling trend on all five tasks: the more robots, the higher the task success rate. In contrast, the HG-DAgger approach, which relies on human teaching, showed only limited scalability, and the collection efficiency of human teaching became the bottleneck limiting its further scaling and breakthrough in success rate.

In the most difficult phone packaging task, the number of robots participating in FERT training expanded from 16 to 96, and the test success rate rose from 86.7% to 98.3%. Within the same real-world interaction duration, a larger-scale robot swarm directly translated into a higher task success rate. After deployment, robots can grow together

CEL collective experience learning provides a scalable direction worth further exploration for embodied intelligence: break away from the traditional route that relies heavily on human demonstrations, and let the robot swarm autonomously accumulate and learn from one another in real operations.

The successful experiences and failure cases encountered by robots under real working conditions are all retained as public assets of the entire cluster. The pitfalls one robot has encountered can be learned from by the whole cluster; effective operations figured out by a single robot can be reused by all robots. When every robot can contribute experience, quantity itself becomes a new dimension of capability growth.

Source: contributed article

Images in this article are from: 123RF licensed image library Infringement report Return to sharing Related articles All comments