Training Humanoids in World Models via Differentiable Heightmap Lookups
breadli428 · x · 2026-11-04
ETH's leggedrobotics group open-sourced Terrain-Grounded Imagination (TGI), which trains a Unitree G1 humanoid to walk over stairs, gaps and stepping stones entirely through a learned world model.
Key idea: the world model learns only the robot's dynamics; terrain, rewards and perception are read from the simulator's heightmap via cheap differentiable lookups at every imagined step, so gaps and edges stay exact instead of being blurred by learned approximation errors.
- Code released under MIT license
- Paper in preparation
- Training happens fully inside the world model, no real terrain interaction needed
More from Embodied
- Gus Robotics watches its Microduck robot production line run for the first time in Shenzhen — mishig25 · 2026-11-05
- Smartglasses CV tells flower pickers what to harvest, cutting seasonal training to zero — nptacek · 2026-11-05
- Open-source universal manipulator: $1,500 BOM, robot-as-a-service at $2,000/month — kamathsblog · 2026-11-05
- Walkie One launches as the world's first AI walkie-talkie for screenless AI access — Scobleizer · 2026-11-04
- SimEX lets coding agents learn robot skills in sim with ~10 min of real robot time — micoolcho · 2026-11-04
- DISCO, CoRL 2026 best-paper finalist: distilled tree search over continuous actions — davheld · 2026-11-04