VR mocap pipeline captures 1,185 human trajectories to teach Unitree G1 to navigate clutter
Scobleizer · x · 2026-09-22
Scobleizer spotlights Void's MTC work: instead of expensive physical obstacle courses, humans are placed in procedurally generated VR clutter scaled to the Unitree G1's body, yielding 1,185 trajectories across 327 environments (32+ hours) of natural invented movements — side-squeezing, ducking, high-stepping, crawling prone. The tricky part is retargeting: frame-by-frame methods can produce poses that look collision-free per frame while the robot's body sweeps through obstacles between frames. The thesis: backflips are cool, but what matters is getting through grandma's storage room without eating drywall.
More from Embodied
- CoRL results show sim-only training now delivers top real-world robot performance — carlosdponx · 2026-11-12
- Robot fails at folding a towel, reassuring roboticists frontier models won't eat robotics yet — chris_j_paxton · 2026-11-12
- Nuro podcast: manufacturing, charging, and regulation gate robotaxi mass adoption — R_D · 2026-11-12
- Rodeo Robotics demos robot that changes your car tire — ycombinator · 2026-11-12
- Cybercab rear-ended by human driver, researcher quips: the problem is the humans — chris_j_paxton · 2026-11-12
- HT-Policy Uses Heavy-Tailed Student-t Regression to Close Robot Policy Gap — chris_j_paxton · 2026-11-12