Quadrupedal World Model accepted at CoRL 2026: zero-shot cross-embodiment locomotion without fine-tuning
breadli428 · x · 2026-10-10
Quadrupedal World Model (QWM) argues a robot's morphology should be routed through learned dynamics, not fed straight to a policy. A generative dynamics model conditioned on scale-invariant physical features — via a physical morphology encoder, adaptive reward normalizer, and latent morphology conditioning — trains policies entirely in imagination. Evaluated on real ANYmal-D and Unitree Go1 (held out), QWM transfers locomotion across embodiments with zero fine-tuning, adaptation, or warm-up, and zero-shot generalizes to unseen morphologies vs. model-free baselines and morphology-expert policies. Accepted to CoRL 2026, Austin.
More from Embodied
- Robots now get up with ease: big upgrades in balance and control — cixliv · 2026-10-10
- Scoble at Tesla Gigafactory: Workers Wear VR-Like Rigs to Train Optimus, 20,000 Already There — Scobleizer · 2026-10-10
- SJTU's UVTA teaches robot dexterous manipulation from 1,000 human tactile demos per task — siyuanhuang95 · 2026-10-10
- SpaceX building industrial base to make up to 1,000 Starships a year, with Florida GigaBay 11x bigger than current Megabay — XFreeze · 2026-10-10
- 20-minute explainer breaks down how Tesla trains FSD: 8 cameras, 36fps, 2B signals to 2 outputs — PTrubey · 2026-10-10
- Real Steel IRL: Humanoid Robots Noisy Boy and Raider-72 Box Each Other — zealcaiden · 2026-10-10