D4orm: Multi-Robot Trajectories with
Dynamics-aware Diffusion Denoised Deformations

IROS 2025
Yuhao Zhang¹, Keisuke Okumura^1,2, Heedo Woo¹, Ajay Shankar¹, Amanda Prorok¹

¹Prorok Lab, University of Cambridge    ²AIST Japan   

This repository includes code for D4orm, based on the model-based diffusion.

D4orm is an optimization framework for generating kinodynamically feasible and collision-free multi-robot trajectories that exploits an incremental denoising scheme in diffusion models.

Installation

To install the required packages, run the following command:

pip install -e .

Run D4orm

To run multi-robot trajectory planning with D4orm, use the following command:

python d4orm/planners/multi_planner.py

Command Line Arguments

Argument	Default	Description
--env_name	multi2dholo	Specifies the environment, including multi2d (Differential Drive), multi2dholo (2D Holonomic), multi3dholo (3D Holonomic).
--Nagent	8	Number of agents.
--Ndiffuse	100	Number of diffusion/denoising steps in one single denoising optimization process.
--Niteration	30	Number of maximum iterations.
--Nsample	2048	Number of samples used in each denoising step.
--Hsample	100	Time horizon (number of timesteps) for the planned trajectories.
--dt_factor	1	Time step scaling factor applied to dt=0.1.
--print_info	(flag)	Flag to print detailed runtime information.
--save_images	(flag)	Flag to save visualization images of planned trajectories.

Modify the arguments to choose a different configuration. Some results will look like this:

Differential Drive	2D Holonomic

Run Baseline Methods

To run baseline methods, use the following command:

python d4orm/planners/path_integral.py

Command Line Arguments

Argument	Default	Description
--env_name	multi2dholo	Specifies the environment, including multi2d (Differential Drive), multi2dholo (2D Holonomic), multi3dholo (3D Holonomic).
--method	mppi	Specifies the method, including mppi and cem.
--Nagent	8	Number of agents.
--Nsample	2048	Number of samples used in each step.
--Hsample	100	Time horizon (number of timesteps) for the planned trajectories.
--save_images	(flag)	Flag to save visualization images of planned trajectories.

Citation

If you find this work to be useful in your research, please consider citing:

@article{zhang2025d4orm,
  title={D4orm: Multi-Robot Trajectories with Dynamics-aware Diffusion Denoised Deformations},
  author={Zhang, Yuhao and Okumura, Keisuke and Woo, Heedo and Shankar, Ajay and Prorok, Amanda},
  journal={arXiv preprint arXiv:2503.12204},
  year={2025}
}