ReDiG: Reinforced Diffusion on Graphs for Decentralized Coordinated Multi-Robot Navigation with Smooth Formation Adaptation

ICLR 2026 Conference Submission20889 Authors

Keywords: Decentralized multi-robot systems, Diffusion model, Reinforcement Learning, Graph Learning

TL;DR: We propose Reinforced Diffusion on Graphs (ReDiG), which unifies graph learning, diffusion models, and online reinforcement learning to enable decentralized coordinated multi-robot navigation with smooth formation adaptation.

Abstract:

Coordinated navigation is a fundamental capability for multi-robot teams to traverse complex unstructured environments.
During navigation, robots are often required to maintain mission-specific formations, such as wedge formations for enhanced visibility and area coverage.
However, rigid formations can hinder navigation in challenging scenarios like narrow corridors, which demand formation adaptation.
Reinforcement learning (RL) is commonly used for coordinated multi-robot navigation due to its ability to learn through interaction with the environment.
However, its step-wise decision-making process often results in jerky motion.
In contrast, diffusion models generate smoother trajectories through probabilistic denoising, but rely heavily on high-quality demonstrations.
Collecting such demonstrations is challenging in multi-robot systems due to the coordination and synchronization required among individual robots.
To address these issues, we introduce a novel method named Reinforced Diffusion on Graphs (ReDiG) to enable
decentralized coordinated multi-robot navigation with smooth formation adaptation.
Under a unified learning paradigm, ReDiG integrates:
(1) graph learning for decentralized coordination to enable formation adaptation,
(2) diffusion models for generating smooth individual robot trajectories, and
(3) online RL to refine noisy demonstrations by leveraging feedback from environment interaction, which enables robot synchronization and guides effective diffusion training.
We evaluate ReDiG through extensive experiments in both indoor and outdoor environments using physical robot teams and robotics simulations.
Experimental results show that ReDiG enables smooth formation adaptation and achieves state-of-the-art performance in coordinated multi-robot navigation within complex environments.
More details are available on the project website: https://anonymous23885.github.io/ReDiG

Primary Area: applications to robotics, autonomy, planning

Submission Number: 20889