Joint Model-based Model-free Diffusion for Planning with Constraints

Wonsuhk Jung*1, Utkarsh Mishra*1, Nadun Ranawaka1,
Yongxin Chen†1, Danfei Xu†1, Shreyas Kousik†1,
1Georgia Institute of Technology
* Equal authorship, † Equal advising,
The link is in preparation
Paper arXiv Video Code

[Need changed] How do we bridge the gap between what robots can do and what they can safely do ? We introduce RAIL : a safety-filter framework for Imitation Learning (IL) policy leveraging reachability analysis to ensure robot safety while maximizing robot capability.

Abstract

Model-free diffusion planners have shown great promise for robot motion planning, but practical robotic systems often require combining them with model-based optimization modules to enforce constraints, such as safety. Naively integrating these modules presents compatibility challenges when diffusion's multi-modal outputs behave adversarially to optimization-based modules. To address this, we introduce Joint Model-based Model-free Diffusion (JM2D), a novel generative modeling framework. JM2D formulates module integration as a joint sampling problem to maximize compatibility via an interaction potential, without additional training. Using importance sampling, JM2D guides modules outputs based only on evaluations of the interaction potential, thus handling non-differentiable objectives commonly arising from non-convex optimization modules. We evaluate JM2D via application to aligning diffusion planners with safety modules on offline RL and robot manipulation. JM2D significantly improves task performance compared to conventional safety filters without sacrificing safety. Further, we show that conditional generation is a special case of JM2D and elucidate key design choices by comparing with SOTA gradient-based and projection-based diffusion planners.

Video

How can we generate diffusion-based plans with feasible safety correction?

Model-based diffusion planners can guarantee constraint satisfaction, but they often produce overly conservative plans failing to satisfy goal-completion objectives.

On the other hand, model-free diffusion planners can generate high-quality plans to achieve goal-completion objectives, but they often violate constraints, leading to unsafe behaviors.

How can we mutually align model-free diffusion planner and model-based optimization module?

Motivation figure showing the gap between model-free diffusion planner and model-based optimization module.

Joint Model-based Model-free Diffusion (JM2D)

A unified diffusion framework for planning and optimization

Key Idea: Recast a planning-and-optimization into a sampling from joint distribution defined by Interaction Potential. In robotics, this interaction potential is usually non-differentiable!

Key Idea: We use Monte-Carlo approximation of score of joint distribution that only needs 0-th order information.

What can JM2D do?

(1) JM2D increase system performance under strict safety guarantee

Text 1

Text 2

Title 1

Text 3

(2) JM2D generalizes Conditional Generation and Model-based Diffusion

Text 1

Text 2

Title 1

Text 3

(3) JM2D generalizes to Joint Model-based Model-free Generation (JM2G)

BibTeX

@inproceedings{
    jung2025joint,
    title={Joint Model-based Model-free Diffusion for Planning with Constraints},
    author={Wonsuhk Jung and Utkarsh Aashu Mishra and Nadun Ranawaka Arachchige and Yongxin Chen and Danfei Xu and Shreyas Kousik},
    booktitle={9th Annual Conference on Robot Learning},
    year={2025},
    url={https://openreview.net/forum?id=E9t1ekt6W9}
}