When to imagine
How should an agent decide whether to run a simulation, extend a rollout, or gather a new observation?
- Adaptive model queries
- Uncertainty and useful representations
- Value of computation and information
Proposed for CVPR 2027First edition
Acting with
World Models
When to imagine, how to act,
and how to recover
We study how agents use world models for policy learning, planning, and recovery.
Prediction into action
Compare possible futures
An illustrative agent decision process
The agent considers several rollouts before choosing an action.
June 20–21, 2027
Seattle, USA
Half day · In person
Workshop acceptance and the exact date and time are pending.
1 / THE QUESTIONS
How agents use world models
Before pouring water, a robot might compare two-second and ten-second predictions of its motion. Which helps it choose an action, and what should it do if the cup moves?
AWM focuses on how agents use learned world models. Topics include learning policies from simulated experience, deciding when to query a model, and responding to prediction errors.
Talks and case studies will examine the agent’s decisions alongside predictive accuracy and the cost of simulation.
How to act
We welcome work on using imagined experience to train policies, evaluate candidate actions, and support planning.
- Learning from imagined experience
- Model-based planning and search
- Video-to-action and vision-language-action (VLA) systems
How to recover
When predictions fail, an agent may need to observe again, replan, or update its beliefs, model, or policy. We are interested in how it chooses among these responses.
- Failure diagnosis and monitoring
- Re-observation and replanning
- Model and policy adaptation
2 / INVITED PERSPECTIVES
Invited speakers
Planned invited speakers. Talk titles and final program details will be announced.
3 / THE PROGRAM
Preliminary program
We plan a four-hour session with six 20-minute invited talks, two paired case discussions, posters, and a 30-minute debate.
Debate
Should world-model rollouts be the default for long-horizon embodied tasks?
The discussion will consider latency, prediction errors, and when to take another observation.| ELAPSED TIME | SESSION |
|---|
Times are relative to the assigned workshop slot. The final date, timing, and session details will follow CVPR decisions. Each 20-minute invited-talk slot includes a 15-minute presentation and 5 minutes of Q&A.
Paired research cases
The first session compares agent strategies using the same world model; the second examines responses to prediction failures. Each session includes two six-minute presentations and eight minutes of discussion.
The AWM card
Speakers will be asked to end with a slide summarizing the task, agent decision, result, resource budget, and a limitation. We plan to publish these slides as a casebook with their permission.
4 / CONTRIBUTE TO AWM
Call for Papers
Preliminary · submissions are not yet open
We welcome methods, analyses, simulation studies, and negative results involving video, latent, or structured world models.
Original research papers
We invite original research papers. We plan double-blind review and publication of accepted papers in the CVPR workshop proceedings, subject to workshop acceptance and conference requirements.
PROPOSED DEADLINEMarch 52027
Submission portal not yet open.
Short papers and research summaries
Short reports, research summaries, paired comparisons, and position papers for presentation without a proceedings entry. Previously published or concurrently submitted work may be eligible, subject to the policies of all relevant venues.
PROPOSED DEADLINEApril 152027
Submission portal not yet open.
The final call for papers, policies, deadlines, and submission link will be published after workshop acceptance.
5 / THE TEAM
Organizers
Researchers working across visual generation, robot learning, physical simulation, and multimodal agents.
Program committee
Confirmed members