CoRL 2026 half-day in-person workshop
A working session on what robotics world models should represent, how they should be evaluated, and how they can be used safely inside robot learning and control loops.
Should models predict pixels, states, action-conditioned futures, latent dynamics, task progress, contact events, or a hybrid? How should actions, proprioception, force, language goals, and task context be grounded across embodiments?
How can world models support short-horizon correction and long-horizon task progress without unstable autoregressive error? When should a model signal completion, recovery, or uncertainty high enough to stop planning?
What demonstrates control utility rather than visual plausibility? Candidate criteria include task success under model-predictive control, 3D and physical consistency, contact prediction, model-policy feedback loops, and human judgments of task progress.
Which datasets, benchmarks, simulation rollouts, policy-in-the-loop protocols, and reporting standards will let the community compare robotics world models across manipulation, dexterous control, manufacturing, and other high-precision settings?
The program combines research talks, contributed spotlights, posters, and a short non-research spotlights and demos session.
We will invite full workshop papers of up to 8 pages and short submissions of 2-4 pages. Short submissions may focus on proof-of-concept results, benchmark proposals, system descriptions, demos, datasets, code, blog posts, negative results, or position papers.
Submissions will use the CoRL workshop format and will be managed through OpenReview with a double-blind review process.
Shao-Hua Sun
National Taiwan University
Kuang-Huei Lee
Google DeepMind
Homanga Bharadhwaj
Johns Hopkins University
Zeyi Liu
Stanford University
Catherine Weaver
Scale AI
Ruoshi Liu
Amazon FAR / University of Maryland
Sean Kirmani
OpenAI
Yu-Wei Chao
NVIDIA
Pepijn Kooijmans
Hugging Face, LeRobot
Jiajun Wu
Stanford University