Simulation physics gaps teach robots tricks that fail in the real world
binarybits · x · 2026-09-04
Citing Kai's new piece on robot data, the author explains a key reason training robot manipulation in simulation doesn't transfer well: because simulation can't perfectly model real physics, models learn exploits that work only in the simulated environment and fail in the real world — a vivid illustration of the sim2real gap.
More from Embodied
- $14k Chinese Cars Now Ship With LiDAR, No Subscription — and BYD Covers Crash Losses — aronchick · 2026-09-04
- Losing a badminton match to a robot: a surprising 2026 milestone — lukas_m_ziegler · 2026-09-04
- Musk: Cybercab interior is a comfortable lounge on wheels with TV and epic sound — elonmusk · 2026-09-04
- Musk shows off driverless Cybercab with no steering wheel or pedals in Austin — elonmusk · 2026-09-04
- Aric Recker secures land for new factory to mass-produce purpose-built machines — MatthewChang · 2026-09-04
- How robot startups scrape for data: free cleanings, exoskeletons, sim limits — binarybits · 2026-09-04