Junwoo Chang* , Hyunwoo Ryu* ,Jiwoo Kim , Soochul Yoo,Jongeun Choi , Joohwan Seo, Nikhil Prakash, Roberto Horowitz
(*: equal contribution)
Neurips 2023 Workshop on Diffusion Models
Abstract
Diffusion models have risen as a powerful tool in robotics due to their flexibility and multi-modality. While some of these methods effectively address complex problems, they often depend heavily on inference-time obstacle detection and require additional equipment. Addressing these challenges, we present a method that, during inference time, simultaneously generates only reachable goals and plans motions that avoid obstacles, all from a single visual input. Central to our approach is the novel use of a collision-avoiding diffusion kernel for training. Through evaluations against behavior-cloning and classical diffusion models, our framework has proven its robustness. It is particularly effective in multi-modal environments, navigating toward goals and avoiding unreachable ones blocked by obstacles, while ensuring collision avoidance.
Collision-avoiding diffusion kernel:
We introduce a novel collision-avoiding diffusion kernel for collision-free motion planning, inspired by heat conduction.
The kernel treats obstacles (black rectangles) as thermal insulators, thereby ensuring our trained model generates collision-free trajectories.
Goal distribution
Collision-avoiding diffusion kernel
Gaussian diffusion kernel
Beginning from the initial goal distribution,
the collision-avoiding diffusion kernel propagates avoiding any obstacles,
whereas the Gaussian diffusion kernel invades through obstacles
How it works
We leverage the heat-inspired diffusion kernel to train our method to jointly generate goal configurations and collision-free trajectories.
Our method "denoises" the agent's state toward the goal distribution.
Assume that each red dot is an "agent (robot)", green apples are the "goal positions", and black rectangles are the "obstacles".
Single goal
Two goals
Three goals
Demos
All experiments are conducted to plan the sample positions of the agent (red dots) toward the goals (apples) while avoiding obstacles (black rectangles).
U map (Experiment) :
Multiple goals
(all reachable)
Multiple goals
(one unreachable)
(This video is not from our method)
The agents (red dots) get stuck in the obstacles (black rectangles),
indicating that if no stationary dots remain, no collisions are occuring
Other maps:
Multiple goals
(all reachable)
Multiple goals
(one unreachable)
BibTeX
@article{chang2023denoising,
title={Denoising Heat-inspired Diffusion with Insulators for
Collision Free Motion Planning},
author={Chang, Junwoo and Ryu, Hyunwoo and Kim, Jiwoo and Yoo, Soochul and Seo, Joohwan and Prakash, Nikhil and Choi, Jongeun and Horowitz, Roberto},
journal={arXiv preprint arXiv:2310.12609},
year={2023}
}