Learning from Fleet Experience: A Collective Learning Paradigm for Robots

Authors: Chao Yu, Shu’ang Yu, Hao Lin, Hongzhi Zang, Yuxiang Yan, Chenxi Jiang, Zhen Guo, Chuankang Li, Yihao Chen, Xin Xu, Xu Fu, Jiakai Zhou, Liangzhi Shi, Hongyi Ren, Yiwei Liu, Jing Wu, Yuanqing Wang, Chen Wu, Bo Dai, Chunyu Yang, Hongliang Huang, Shuoyu Yue, Weiqi Yin, Yinuo Chen, Ruihua Chen, Chunyang Zhu, Han Xiao, Yanghao Li, Ruipu Wu, Zhikai Cai, Xiu Chen, Xu Tan, Yuchen Yan, Jialin Zhang, Wen Zeng, Yiming Zhou, Zixiao Huang, Guanghui Ren, Quanlu Zhang, Boxun Li, Maoqing Yao, Yu Wang · arXiv 2026

Licences

Open Access
Paper: CC BY 4.0

Everything below draws on the full paper under its CC BY 4.0 licence; its figure is credited to the authors.

TL;DR

Deployed robots share successful runs and failures decide which ones to retrain on. With up to 96 AgiBot G2 robots and no new expert actions, five tasks reach 93.3–99.2% success.

Why it matters

After initialization, the only human input during FERT is a binary success label per episode and environment resets; the actions used for learning come from the robots themselves. All policies were evaluated on two robots that contributed no training data, so the reported gains transfer across robot instances of the same platform.

Problem

Improving a deployed robot policy usually needs teleoperators who watch execution, spot failures and take over to supply corrective actions. That supervision is hard to scale as a fleet grows, even though a larger fleet produces more execution experience that could be used for learning.

Proposed solution

Collective experience learning (CEL) pools experience from all deployed robots to update one shared policy. Its algorithm, FERT, keeps successful trajectories as action targets and uses the roughly three seconds before each failure as a query, matched by slope-constrained subsequence dynamic time warping on frozen visual features plus joint state, to raise the sampling priority of related successful segments during flow-matching fine-tuning of π0. The fault-tolerant asynchronous training system (FATS), with a SwarmLink data plane and a HotSwarm control plane, keeps collection, cloud inference and training running while individual robots fail and are restarted in place.

Key insights

Failed runs are useful as retrieval queries even though their actions are never imitated: removing failure-guided retrieval cut Phone Packaging success (16 robots) from 86.7% to 64.2%.
With collection time per robot held fixed, FERT success generally rose with fleet size, while human-gated HG-DAgger saturated (Phone Charging, Phone Packaging) or declined (Stool Assembly) as fleets grew; at 8 robots FERT could fall below its starting policy.
Extending each retrieved segment by 30 frames on each side matters: without the extension, success fell from 86.7% to 57.5%.
Local fault recovery keeps the fleet productive: 30 fault incidents on 22 of 96 robots removed 8.3 of 660 robot-hours (1.3%), and after a median fault event the fleet completed 99% as many episodes as before it.
In a trace replay, synchronous collection rounds would have gathered only 33% of the run's episodes, because every round waits for the slowest robot's episode or pause.

Main results

Mean success over five tasks rose from 36.3% (SFT) to 81.0% (multi-robot HG-DAgger) to 97.3% (FERT), an 86.0% relative reduction in failure rate; 120 trials per task on held-out robots.
Phone Packaging with 96 robots: 79.2% → 98.3%; Phone Charging: 60.0% → 97.5% (93.8% relative failure reduction).
FERT beat RLT and RECAP on all five tasks by 12.3 and 11.5 percentage points on average, and exceeded six human teleoperators by 16.7 points on average.
Phone Packaging success grew from 86.7% with 16 training robots to 98.3% with 96.
The 96-robot run collected 9,834 episodes; outside the lunch break about 65 robots executed at once and the fleet completed 24 episodes per minute; robots resumed a median 12 s after an incident ended.
Over the 30 Gbps edge-cloud link, 95.18% of observation transfers finished within 20 ms and 83.15% of action-chunk replies took 200–250 ms.

Limitations

The authors state the study evaluates task-specific policies on five tasks using a single robot platform; broader task sets, heterogeneous robot platforms and adaptation over longer periods are left to future work.
Humans still label every episode's outcome and reset environments; a learned reward model is mentioned as an option but not used.
FERT starts from a policy already trained with expert demonstrations (SFT) and human interventions (multi-robot HG-DAgger).
At small fleet sizes FERT can end below its initial policy, for example with 8 robots on Stool Assembly and Phone Packaging.
The local improvement guarantee (Theorem 1) holds only under the stated replay and retrieval-alignment assumptions and a sufficiently small step size.

Field impact

On Hugging Face daily papers it drew 112 upvotes, in the top 5% of robotics papers there over the last six months. The paper gives a project page (cel-robot-learning.github.io); its text does not state a code or data release. FATS is built on the RLinf framework.

Abstract

Experience generated during deployment offers an opportunity for robots to improve policy reliability through learning from their own executions. However, learning through human corrections requires teleoperators to monitor execution and intervene when policies fail, a demand that is difficult to meet as robot fleets grow. We introduce collective experience learning (CEL), a learning paradigm in which deployed robots share their execution experience to improve a common policy. We realize CEL through failure-guided experience retrieval fine-tuning (FERT), which uses failure contexts from robot executions to retrieve relevant trajectory segments from the fleet's shared successful experience and uses their actions as supervision for policy updates, without collecting additional expert demonstrations or corrective actions. A fault-tolerant asynchronous training system (FATS) coordinates experience collection, policy optimization and policy distribution despite local interruptions within the robot fleet. Across five real-world manipulation tasks, the resulting policies achieve success rates of 93.3-99.2%, evaluated over 120 trials per task on robots excluded from training data collection. With collection time per robot held fixed, larger fleets provide more experience and FERT generally achieves higher success rates across the evaluated sizes. On multi-stage phone packaging, learning with 96 robots increases success from 79.2% to 98.3%. These results support collective experience as a practical basis for scalable robot learning and a path towards improving policy reliability during deployment without collecting additional expert corrective actions.

Open Source Robotics