Quadrupedal World Model Accepted to CoRL 2026: Zero-Shot Cross-Embodiment Locomotion

Quadrupedal World Model (QWM), a study led by Mohamad Danesh in collaboration between McGill, Mila, Université de Montréal, and the leggedrobotics team at ETH Zurich, has been accepted to CoRL 2026 and will be presented in Austin, USA. The paper's core claim: a robot's morphology specifications should not be fed directly to the policy network, but should first be routed through a learned dynamics model, with the policy then extracted within the model's "imagination".

Confirmed

• Motivation: With robot morphology formally specified, there are two ways to build a controller for a family of robots—feed the specifications to a model-free policy, or feed them to a learned dynamics model and extract the policy in imagination; the authors argue for the latter and build QWM accordingly.
• Method details: A single generative dynamics model conditioned on scale-invariant physical features, comprising three components—a physical morphology encoder, adaptive reward normalization, and latent morphology conditioning; the policy is trained entirely inside the model (in imagination).
• Key capability: A single world model achieves zero-shot transfer across multiple quadruped robots, with no fine-tuning, no warm-up, and no per-embodiment retraining.
• Affiliation: Collaboration between McGill, Mila, Université de Montréal, and ETH; accepted to CoRL 2026, to be presented in Austin, USA.

2026-10-10 ~ 2026-10-10 · 7 related posts

• [source] Quadrupedal World Model accepted at CoRL 2026: zero-shot cross-embodiment locomotion without fine-tuning — breadli428 · 2026-10-10
• [source] QWM thread: why morphology specs should go through a learned dynamics model, not model-free policies — breadli428 · 2026-10-10
• QWM zero-shot transfer to unseen morphologies vs. model-free and morphology-expert baselines — breadli428 · 2026-10-10
• QWM method: morphology encoder, adaptive reward normalizer, and latent morphology conditioning — breadli428 · 2026-10-10
• [source] QWM on real ANYmal-D and Unitree Go1: cross-embodiment transfer with no fine-tuning or warm-up — breadli428 · 2026-10-10
• CoRL 2026 paper: robot morphology should be routed through learned dynamics, not fed to policy — breadli428 · 2026-10-10
• QWM: one world model zero-shot transfers locomotion across quadruped robots at CoRL — breadli428 · 2026-10-10