Call for Papers

Overview

The World Models in Physical AI Workshop at NeurIPS 2026 invites submissions on learned models of physical-world dynamics for Physical AI.

World models predict how an environment evolves, optionally conditioned on an agent's actions, and can be rolled out to imagine future observations and outcomes. For Physical AI systems — including robots, autonomous vehicles, and other agents that perceive and act in the physical world — this capability is central: an agent that can simulate its world can plan, learn from imagined experience, and address situations not represented in its training data.

The workshop will bring together researchers from generative modeling, reinforcement learning, robotics, computer vision, autonomous driving, and simulation to discuss how world models can become actionable, physically grounded, reliable, and deployable.

Topics of Interest

We welcome submissions on all aspects of world models for Physical AI. The topics below are illustrative rather than exhaustive; submissions are not required to fit neatly within any one theme.

Representations and Architectures

Latent versus pixel or video models, JEPA-style embeddings, identifiable latent representations, omnimodal models, 3D dynamics, contact dynamics.

World Models for Action

Model-based reinforcement learning, planning and control, learning in imagination, forward and inverse dynamics, world-action models, action-conditioned prediction.

Generative Simulation

World models as data generators, interactive and closed-loop simulation for robotics and autonomous driving, sim-to-real, real-to-sim, generative simulators and digital twins.

Evaluation

Physical correctness, causal faithfulness, downstream control utility, policy-ranking preservation, robustness, generalization, benchmarks and evaluation protocols.

Scaling and Foundation World Models

Data and compute scaling, large pretrained world models, multimodal data, transfer across embodiments, action spaces and domains.

Safety, Reliability and Broader Impact

Uncertainty, controllability, long-horizon consistency, failure analysis, safe deployment and the broader impact of world models that drive physical systems.

We particularly encourage work that addresses open questions at the intersection of these themes, including pixels versus latents, controllability over long horizons, sim-to-real and real-to-sim transfer, and the scaling behavior of world models for Physical AI. We also welcome relevant interdisciplinary work and research directions not explicitly represented in this list.

Submission Format

Papers may be up to 8 pages, excluding references. All submissions must use the NeurIPS 2026 paper template.

Review is double blind. Submissions must be anonymized: remove author names and affiliations, avoid identifying acknowledgements or funding information, and refer to your own prior work in the third person. Please also avoid links or supplementary materials that reveal author identity.

Submissions will be managed through OpenReview. Submit your paper at the workshop's OpenReview submission portal.

Use of Agents and Large Language Models

This workshop follows the NeurIPS 2026 Main Track Handbook and the NeurIPS 2026 Workshops Guidance.

Authors may use tools to assist in preparing a submission, but they remain fully responsible for the entire content of the paper and must ensure that all text, figures, references, claims and results are correct, original and scientifically responsible. Routine spell checking, grammar suggestions, editing assistance and basic code assistance do not need to be documented. If agents or LLMs are an important, original or non-standard component of the research method, their use must be described in the experimental setup or an equivalent section. Agents and LLMs may not be listed as authors. Prompt injections and other attempts to manipulate the reviewing process are strictly prohibited.

In addition to these conference-wide requirements, the workshop strongly discourages using LLMs or other automated systems to generate a submission's scientific contribution — including its core ideas, claims, methods, experiments, analysis, figures or references — without substantive, author-led intellectual engagement and careful verification.

Consistent with the NeurIPS workshop guidance, reviewers will not be permitted to use LLMs or agents when reviewing workshop submissions.

Review and Publication Policy

• Each submission will be peer reviewed through OpenReview under a double-blind process; submissions must be anonymized.
• At least one author of every submission must agree to serve as a reviewer. Others can volunteer via the reviewer sign-up form.
• Work already published or presented at the NeurIPS 2026 main conference is not eligible for submission to the workshop.
• Accepted submissions are non-archival, allowing authors to submit the work to an archival conference or journal in the future, subject to the policies of that venue.
• Details concerning the public posting of accepted submissions on OpenReview and the workshop website will be announced before submissions open.

Conflicts of interest will be managed using OpenReview's conflict-of-interest tools. Organizers will not review their own conflicted submissions or work from authors at their own organizations. People with a conflict of interest with an organizer, including their current students or postdoctoral researchers, may not submit.

Presentation

Accepted submissions will be presented during the one-day, in-person workshop in Sydney, Australia. All accepted papers will be invited to the poster session, and selected papers will also be invited for contributed spotlight talks.

The workshop program will include invited talks, contributed presentations, poster sessions, a panel on evaluating world models for Physical AI, and a debate on pixels versus latent representations.

Important Dates

All deadlines are Anywhere on Earth (AoE) unless otherwise stated.

Inclusion

We welcome participation from researchers across disciplines, sectors, career stages, geographic regions and backgrounds. We particularly encourage submissions and participation from groups and regions that have historically been underrepresented in machine learning, robotics and Physical AI.