Tsinghua, AgiBot, Infinigence propose CEL: 96-robot fleet learning lifts success 79.2% to 98.3%
jiqizhixin · x · 2026-11-15
Tsinghua University, AgiBot, and Infinigence AI introduce Collective Experience Learning (CEL) in an arXiv technical report, demonstrated with a training fleet of up to 96 robots.
- Bottleneck: More robots yield more experience, but expert operators cannot provide corrective actions at the same scale.
- Mechanism: Failure-guided Experience Retrieval Fine-Tuning searches a shared pool of successful trajectories for segments resembling recent failures and prioritizes them for policy updates; an asynchronous, fault-tolerant system keeps experience collection and training running through local interruptions.
- Payoff: On phone packaging, success rises from 79.2% to 98.3% over 120 trials per policy on two same-model robots outside the training fleet.
More from Embodied
- OOMWOO: open-source DIY robot vacuum hits 13.1k GitHub stars with Pi, LiDAR and ROS2 — Technovangelist · 2026-11-15
- Fixing EMI from large servos: foil shielding or just shorten the wires — philfung · 2026-11-15
- AirPods Used for Full-Body Tracking in VRChat Demo, Blowing Minds — Scobleizer · 2026-11-15
- TAU teleoperated cleaning robot caught struggling on the job — Distinct-Question-16 · 2026-11-15
- China unveils autonomous electric passenger jet: 19 seats, 300 km/h, 400 km range — Scobleizer · 2026-11-15
- Robot whole-body controller training drops from 128 GPUs × 1 week to 1 GPU × 3 minutes — StewartalsopIII · 2026-11-15