Lightwheel trains humanoid repair policies on 8+ minute real-world sequences

CyberRobooo · x · 2026-11-18

Lightwheel's RoboFinals RL-Env builds Newton-native simulation environments from EgoSuite first-person captures of real repair jobs, with calibrated hose and contact physics, training policies on multi-stage sequences running 8+ minutes without resets. Post-RL policies complete steps the pre-RL version cannot start, though real shop-floor performance is still unproven.

Related event: Lightwheel Releases RoboFinals RL Environment Trained on Real Repair Data(3 posts)→

Original post →

More from Embodied

Embodied channel →