Overview
Recent advances in learned world models have shown impressive ability to generate visually plausible futures from images and videos, but visual prediction alone is insufficient for interaction-rich robot manipulation. A model that predicts pixel flow may still miss contact dynamics, frictional contact constraints, and underactuated physics needed for reliable control. On the other hand, modern analytical physics simulators have historically been relied on to show the great success in modeling the underlying contact physics and enabling robotic control. Thus, a natural question to ask is How can the expertise of modern simulation development enhance world model learning for better robotic control?
We aim to bridge the historically disparate communities of physics simulation and deep representation-based world modeling. This half-day workshop will be discussing around the following four questions.
Research Questions
This workshop will focus on the following specific research questions at the intersection of world models, physics simulation, and control for manipulation:
- Q1 — Why do robots need predictive models? Imitation learning and model-free RL have achieved impressive manipulation results, but deployment still requires reasoning about physics understanding, long-horizon consequences, safety, and failure recovery. What roles should predictive models—including physics simulators and learned world models—play? When are they essential, and what capabilities remain difficult to obtain from model-free learning alone?
- Q2 — What makes a world model useful for robotic manipulation? Visually plausible prediction is not enough for control. A useful world model should capture the aspects of physics that affect action selection, such as contact transitions. Which physical quantities can be learned from passive video, which require interaction or additional modalities, and what benchmarks can verify that a model is control-useful rather than only visually convincing?
- Q3 — How should predictive models be built? Predictive models span analytical simulators, video world models, structured neural models, and hybrid neural–physics systems. The community has not yet converged on a specific method. What should be scaled purely from data, and what physical structure should be built in? In light of the bitter lesson, when do physics priors help rather than restrict learning, and how can simulators, interaction data, and learned representations be combined for a more effective modeling paradigm?
- Q4 — How should robotic manipulation handle imperfect predictive models? Eventually, all simulators and learned world models are biased approximations of reality. Rather than pursuing perfect fidelity, how can manipulation handle imperfect predictive models in policy acquisition, data generation, and evaluation? How should control policies reason about model uncertainty, compounding error, and exploitable artifacts without failing in the real world?
Invited Speakers & Panelists
Speaker list is being finalized. Status is shown for each invitee.
Yuval Tassa
Google DeepMindSpeaker & Panelist · Confirmed
Erwin Coumans
NVIDIASpeaker & Panelist · Confirmed
Additional Panelists
Event Schedule
The workshop is a half-day, in-person event. Absolute start times will be announced once the CoRL program is finalized; the planned structure is below.
| 10 min | Opening Remarks |
| 2 hours | Invited Speaker Presentations (four talks, ~25 min each) |
| 30 min | Contributed Spotlight Presentations |
| 30 min | Coffee Break & Poster and Demo Session |
| 1 hour | Panel Discussion with Live Q&A |
| 10 min | Awards & Closing Remarks |
Accepted Papers
Accepted papers will be announced here following the review process. See the Call for Papers for submission details.
Call for Papers
Submission Portal: OpenReview (link to be announced)
We cordially invite paper submissions relevant to the following (non-exhaustive) topics:
- Physical Fidelity and Evaluation of World Models
- Injecting Structural Physics Priors into Learned Representations
- Integrating Analytical Simulation with Neural World Models
- World Models for Planning, Control, and Decision Making
- Learning Physical Dynamics from Video and Multimodal Data
- State-Action Representations and Controllable Abstractions
- System Identification, Residual Dynamics, and Real-to-Sim
- Sim-to-Real Transfer and Long-Horizon Reasoning with World Models
Submission Guidelines
- Deadline: To be announced.
- Page Limit: We welcome submissions of up to 8 pages (excluding references) describing completed research, ongoing work, or benchmark datasets relevant to the workshop topics.
- Formatting: Authors are encouraged to use the CoRL 2026 paper template. Templates from related conferences will also be accepted.
- Reviews: Each submission will receive at least two reviews from the organizing committee and external reviewers. Authors of submitted papers are expected to help review other submissions.
- Anonymity: All submissions will be double-anonymized.
- Dual Submission: We welcome submissions that are under review or recently published at relevant venues.
- Non-Archival: This workshop is non-archival. Accepted papers will be posted on the workshop website.
- Presentation: Accepted papers will be presented in an interactive poster session, with the top 3–5 papers selected for oral spotlight presentations. Authors are expected to present in person.
- Best Submission Award: A Best Workshop Submission Award ($500 cash prize, sponsored by DAXO Robotics) will be presented, subject to CoRL policy.
Organizers
Sponsors
