CEL Project Page: Paper and Code for Collective Experience Learning Across 96 Robots
CyberRobooo · x · 2026-10-25
The author shared the project page, demos, paper, and code links (cel-robot-learning.github.io) for CEL — the same work as the earlier tweet.
The page details the evolution of three learning paradigms: SFT relies on human demonstrations, (M)HG-DAgger on expert corrections, while CEL makes the fleet's own successful executions the supervision signal — drastically reducing expert supervision needs. Humans only provide outcome labels and reset environments. Improvements are especially notable on long, multi-stage industrial tasks like phone packaging.
This is the most primary source (paper + code) for readers who want to dig in or reproduce.
More from Embodied
- Nov 7 workshop: control a robot dog hands-free with an ESP32 — StewartalsopIII · 2026-10-26
- Tesla's Texas robotaxi fleet nears 1,000 vehicles, up 78% in October — mitchdeg · 2026-10-26
- Are we evaluating robot dexterity wrong? The case for superhuman tasks — DJiafei · 2026-10-26
- One prompt had AI optimize Gazebo: 68-79% CPU savings and up to 3.6x speedups, identical results — facontidavide · 2026-10-26
- MEMI Reverse-Engineers Robot Memory, Hits 97% on RoboMME (99.4% with Controller) — du_yilun · 2026-10-25
- Musk: Mine lunar silicon to build AI satellites — Earth launches can't scale to 500-1,000 TW a year — elonmusk · 2026-10-25