Diffusion for Robot Learning

Vision

Diffusion has established itself as a dominant learning and generation paradigm, emerging as a powerful alternative to autoregressive schemes. While its impact on image and video synthesis is well-established, its potential to reshape robot learning is vast. We believe the robot learning community is currently entering a more embracing phase for diffusion methods, where they serve as a foundation for both high-fidelity data generation and intelligent control.

This workshop aims to aggregate recent advances in diffusion modeling into trackable robotics capabilities and pinpoint the most promising future directions for the community, standing at the intersection of generative AI and physical intelligence.

Topics of Interest

Diffusion as a Policy

Data-efficient and scalable training for generalizable policies.

Steering Policies

Gradient-based and classifier-free guidance for goal-reaching.

Diffusion for VLA

Integrating diffusion into Vision-Language-Action models.

Dynamics of Denoising

Analyzing the iterative action formation process.

Accelerating Inference

Few-step diffusion, Consistency Models, and Flow Matching.

World Models

Semantic physics knowledge from large video models.

Motion Synthesis

Diffusion-based generation of high-fidelity human/hand poses to solve data scarcity.

Invited Speakers

Yilun Du

Harvard University

Negar Mehr

UC Berkeley

Max Simchowitz

CMU

Ian Manchester

University of Sydney

Workshop Schedule (July 17)

Timeline	Speaker	Title
2:00 PM - 2:10 PM	Organizers	Opening Remarks
2:10 PM - 2:40 PM	Ian Manchester	Structured Diffusion for Robot Learning: Safety, Speed, Generalization, and more…
2:40 PM - 3:10 PM	Yilun Du	Model Based Decision Making with Diffusion Models
3:10 PM - 3:40 PM	Spotlight Authors	Diff4RL Workshop Spotlight Talks
3:40 PM - 4:00 PM	Ran Yu	Sponsor Talk (X Square Robot)
4:00 PM - 4:30 PM	Poster Authors	Poster Session & Coffee Chat
4:30 PM - 5:00 PM	Max Simchowitz	Diffusion: Do we need it in robot learning?
5:00 PM - 5:30 PM	Negar Mehr	Diffusion Models that Robots Can Execute Together
5:30 PM - 6:00 PM	Spotlight Authors	RSS Diffusion Paper Spotlight Talks
6:00 PM - 6:10 PM	Organizers	Closing Remarks & Awarding

Title	Presenter	Time
Exploring the Intrinsic Geometry of Diffusion Models with Constrained Inverse Kinematics	Phone Thiha	3:10 – 3:15
Uncertainty Quantification for Flow-Based Vision-Language-Action Models	Ralf Römer	3:15 – 3:20
Improving Robotic Generalist Policies via Flow Reversal Steering	William Chen	3:20 – 3:25
Start Right, Arrive Right	An Thai Le	3:25 – 3:30
FlowDPG	Kexin Shi	3:30 – 3:35
DAM-VLA	Pankhuri Vanjani	3:35 – 3:40

Title	Presenter	Time
Set-Supervised Diffusion Policy: Learning Action-Chunking Diffusion through Corrections	Zhaoting Li	5:30 – 5:37
SanD-Planner: Sample-Efficient Diffusion Planner in B-Spline Space for Robust Local Navigation	Jingcheng Wang	5:37 – 5:44
HumanFlow: Diffusion-Driven MAV Navigation Among Humans via Tightly-Coupled Motion Tracking, Forecasting, and Control	Simon Schaefer	5:44 – 5:51
TMRL: Diffusion Timestep-Modulated Pretraining Enables Exploration for Efficient Policy Finetuning	Matthew Hong	5:51 – 5:58

Call for Papers

We welcome submissions that explore the intersection of generative modeling and physical intelligence, ranging from theoretical foundations to large-scale robotic deployments.

Submission Guidelines

Paper Length: 4 to 8 pages (excluding references and supplementary material).
Format: Submissions must follow the RSS paper template.
Review Process: Peer review will be single-blind.
Dual Submissions: As a non-archival venue, we welcome submissions of work that is currently under review or has been recently accepted/published at other venues.

Presentation & Awards

All accepted papers will be presented during an on-site poster session. A subset of high-quality submissions will be selected for 5-min spotlight talks. During the closing remarks, we will present two awards, sponsored by X Square Robot: a Best Paper Award ($1,000) and a Best Student Paper Award ($500). Accepted papers will also be made available on the workshop website.