A Debate on Dynamics Learning, Planning, and End-to-End Robot Learning
World models are rapidly becoming a central paradigm in robot learning. Broadly, they are predictive systems that let robots anticipate how the world may change under their actions — spanning dynamics, contact, geometry, and physics-based models, as well as latent predictors, video models, learned simulators, and world–action models. Yet their growing impact has exposed an unresolved tension: should predictive structure be made explicit to support planning, counterfactual reasoning, data efficiency, safety, and generalization — or can it be absorbed implicitly into end-to-end policies that map observations and goals directly to actions?
World models offer reusable structure but may introduce brittleness, hallucinated predictions, modeling errors, slow inference, or unnecessary complexity, while imitation, diffusion, and vision–language–action (VLA) policies suggest some of this structure can be learned from data without exposing a separate model for prediction or planning. This workshop turns that tension into a structured scientific debate — asking not whether world models are necessary in the abstract, but when, why, and in what form they help, where they introduce brittleness, and when end-to-end policies are the better choice. Our goal is to turn the debate into concrete design principles for choosing among explicit world models, implicit predictive policies, and hybrid systems.
The question cuts across reinforcement learning, control, computer vision, simulation, and embodied AI. The workshop is built for a broad audience: newcomers who want a grounded introduction to world models, and experts from academia and industry who want to debate when and why to use them — and confront their limits.
Rather than surveying world-model approaches broadly, the workshop stress-tests six concrete claims shaping the field today. Each is used as a debate motion: invited speakers defend assigned sides, the audience votes before and after, and a community white paper consolidates the outcome.
Visuomotor, diffusion, and VLA models may already embed action-conditioned knowledge about dynamics without exposing it for planning.
Open problem: when is implicit structure sufficient, and when does its absence cause brittleness under unseen objectives, novel objects, longer horizons, distribution shift, or failure recovery?
Learning world models provides an independence between dynamics and task objectives, supporting planning for novel goals and test-time scaling of compute.
Open problem: what characteristics of a downstream task demand an explicit model or make it better suited to be learned from end-to-end training — and how should this be quantified?
A policy need only retain information relevant for action selection, whereas a reusable model demands broader counterfactual coverage and faces challenges in defining a state abstraction.
Open problem: is it easier to scale learning for general-purpose policies, or can one tractable model abstraction support broad reuse and learning from broader data?
Policy evaluation drives policy improvement and is critical for assessing behavior at deployment time, yet real-world evaluation may be costly or dangerous.
Open problem: when does reproducible model-based evaluation preserve real-world policy rankings and improvement directions — or introduce a new failure mode: trusting a model of the world that is itself wrong?
Predictive dynamics can appear as a deploy-time planner, a learned simulator, a training-time data generator, a pre-trained policy representation, a safety monitor, or an evaluation tool.
Open problem: which role fits which regime, given that a model useful for pretraining may fail at online control, and image prediction may not support contact-rich manipulation?
Visual prediction quality is insufficient, and task success alone hides differences in sample efficiency, robustness, and generalization.
Open problem: how do we design comparisons under matched data, compute, and deployment constraints — covering distribution shift, long horizons, contact-rich interaction, failure recovery, and real-world transfer?
Stanford University / Physical Intelligence
Columbia University
Technical University of Darmstadt
Massachusetts Institute of Technology / Rhoda AI
These debate assignments are intended to create a rigorous discussion and do not necessarily represent each participant's personal views on every motion.
NVIDIA
Technical University of Darmstadt
Stanford University / Physical Intelligence
Massachusetts Institute of Technology / Rhoda AI
Columbia University
Half-day program. Invited talks are 20 min + 5 min Q&A. Times are tentative and will be finalized once the CoRL 2026 program is set.
| Time | Duration | Session |
|---|---|---|
| 08:50–09:00 | 10 min | Welcome & opening vote on the six motions |
| 09:00–09:25 | 25 min | Invited talk: Chelsea Finn |
| 09:25–09:50 | 25 min | Invited talk: Yunzhu Li |
| 09:50–10:20 | 30 min | Contributed spotlights |
| 10:20–11:00 | 40 min | Coffee break & poster session |
| 11:00–11:25 | 25 min | Invited talk: Georgia Chalvatzaki |
| 11:25–11:50 | 25 min | Invited talk: Vincent Sitzmann |
| 11:50–12:40 | 50 min | Structured debate & closing vote — moderated by Yunzhu Li |
| 12:40–12:50 | 10 min | Awards & closing remarks |
We solicit two kinds of contributions via OpenReview — standard research papers and short position papers arguing for the relevant topics. We explicitly invite in-progress work, head-to-head comparisons, and negative or surprising results (a paradigm that failed where it was expected to win) — exactly the evidence the debate needs. We only accept submissions that have not been published elsewhere or accepted to the main CoRL 2026 conference.
Both tracks have a 4-page limit, excluding the bibliography and appendices, and use the CoRL template.
Review is double-blind and follows the CoRL main-track reciprocal model: each submission nominates one author to review another, ensuring at least two reviews per paper, with an organizer serving as editor. Accepted work appears as spotlights and posters, with an optional short video. Two targeted awards recognize both empirical and argumentative contributions: Best Research Paper and Best Position Paper.
Topics include: explicit and latent world models for control and planning; end-to-end, diffusion, and VLA policies; hybrid model-based / model-free systems; counterfactual and long-horizon reasoning; uncertainty and safe deployment; and protocols benchmarking explicit vs. implicit approaches under data and compute.
Receives an NVIDIA Jetson Thor + Samsung Galaxy Tab S11 Ultra.
Receives an NVIDIA Jetson Thor.
Prizes generously provided by our sponsors, NVIDIA and Samsung.
All deadlines are at 11:59 PM Pacific Time.
| Submission opens (CfP announced) | September 21, 2026 |
| Paper submission deadline | October 12, 2026 |
| Review deadline | October 23, 2026 |
| Notification to authors | October 26, 2026 |
| Camera-ready deadline | November 6, 2026 |
| CoRL 2026 conference starts | November 9, 2026 |
| Workshop day | November 12, 2026 |
| Optional revised version deadline | December 9, 2026 |
Thank you to NVIDIA and Samsung for supporting the workshop and paper awards.
