REA embodied agent framework hits 95.9% on LIBERO-Pro with 25% fewer control steps
YangYou1991 · x · 2026-11-07
REA is an embodied agent framework that lets VLA and code policies keep acting while the reasoning agent thinks asynchronously — the robot body doesn't wait for its deep-thinking brain.
- 95.9% success across 3,000 LIBERO-Pro trials, up from 94.7% for the strongest single baseline
- 25% fewer control steps per success
- Notably, REA picked policies from past experience just as well as asking an agent live
Paper, code, and demos are publicly available.
More from Embodied
- Cline creator: imitation learning is a local minimum, evolution is the global one — MarcJSchmidt · 2026-11-07
- CMU's Pixel: wheeled humanoid you can teach by wearing a sensorized rig — lukas_m_ziegler · 2026-11-07
- SynTalk: open-source gesture-to-speech with dual camera and flex-glove pipelines — garvgoel2005 · 2026-11-07
- Neura Robotics opens Hangzhou branch, 4NE-1 to weld at HD Hyundai shipyard — CyberRobooo · 2026-11-07
- India's i-Hub Robotics shows welding demo with wheeled humanoid Daksha Gen 2 — CyberRobooo · 2026-11-07
- Shanghai Kepler's humanoid K2 does remote VR-controlled high-altitude welding — CyberRobooo · 2026-11-07