Robots need failure data: learned sims handle rendering, MuJoCo keeps contact honest
chris_j_paxton · x · 2026-11-04
srigopal notes gripper-closure artifacts are expected since contact is where learned sims have least data. His practice: replaying real SO-101 teleop in MuJoCo for eval, with joints tracking within 1°, so you know where the sim is lying; he predicts learned models will handle rendering while physics engines handle contact. chrisjpaxton adds this is exactly why the field must collect failure data, pointing to a recent RoboPapers episode on the topic.
More from Embodied
- Primordiant launches general-purpose robot scientists to run experiments at scale — ycombinator · 2026-11-04
- D&A turns any VLM into a zero-shot robot reward model with SOTA on three benchmarks — KyleMorgenstein · 2026-11-04
- Waymo announces major expansion of its LA robotaxi service today — reed · 2026-11-04
- CMU releases Pixel: a low-cost wheeled humanoid with co-designed UMI data pipeline for mobile manipulation — ankurhandos · 2026-11-04
- Kid uses Rabbit R1 to recall 2010 World Cup results — 'most underrated device of the last five years' — SimonBalmain · 2026-11-04
- Robotics researcher: training neural sims needs ~10x more failure than success data — chris_j_paxton · 2026-11-04