CoDiMAD Diffusion-Based Privileged Distillation for Communication-Free Multi-Robot Coordination

Jul 16, 2026

Decentralized multi-robot coordination under partial observability is especially difficult in communication-free settings, where agents must act from local sensor observations alone. Privileged policy distillation can transfer behavior from a globally informed oracle to sensor-constrained students, but multi-agent ambiguity makes the conditional oracle action distribution inherently multi-modal. CoDiMAD addresses mode collapse in deterministic distillation using a three-stage diffusion-based framework, yielding coherent coordination actions via diffusion reverse sampling and providing theoretical guarantees and improved experimental performance.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages8

Reading Time20 min

Introduction

Decentralized execution under partial observability

Privileged policy distillation

Multi-modal action ambiguity in multi-agent systems

CoDiMAD diffusion-based distillation approach

CoDiMAD: Diffusion-Based Privileged Distillation for Communication-Free Multi-Robot Coordination Jiyue Tao 1 , Shunheng Xin 1 , Tongsheng Shen2 , Dexin Zhao2 ,∗ , and Feitian Zhang 1 ,∗ arXiv :2607 .09587v1 [ cs .RO] 10 Jul 2026 Abstract—Decentralized multi-robot coordination under partial observability remains challenging, especially in communication-free settings where agents must act solely from local sensor observations. Privileged policy distillation provides a promising approach by transferring knowledge from a globally informed oracle to sensor-constrained students. However, in multi-agent systems, the same local observation may correspond to multiple global configurations requiring qualitatively different cooperative actions, making the conditional action distribution inherently multi-modal. Standard deterministic distillation collapses these modes to their mean, often yielding invalid or hesitant actions. To address this issue, we propose CoDiMAD, a three-stage framework that trains a privileged oracle with MAPPO, constructs an offline dataset of localobservation-oracle-action pairs, and distills the oracle into decentralized students parameterized as conditional denoising diffusion probabilistic models. By approximating the conditional oracle-action distribution through the diffusion reverse process, CoDiMAD samples decisive actions from coherent coordination modes rather than averaging across them. Theoretical analysis characterizes the mode-averaging failure of deterministic distillation and the distributional recovery property of diffusionbased distillation. Experiments on three cooperative tasks show that CoDiMAD consistently outperforms direct local MARLand deterministic distillation baselines. The source code will be made publicly available upon acceptance. Index Terms—Multi-robot systems, imitation learning, reinforcement learning, multi-agent diffusion policy. I. INTRODUCTION MULTI-ROBOT coordination underpins a growing range of applications, such as environmental monitoring [1], search and rescue [2], and cooperative manipulation [3], where teams of autonomous robots collectively accomplish objectives beyond individual capabilities. In practical deployments, scalability and communication constraints often necessitate decentralized execution [4], where each agent acts based on its own sensor observations without relying on a central coordinator at runtime. The widely adopted Centralized Training Decentralized Execution (CTDE) paradigm [5]–[7] addresses this by leveraging global information during training, typically through a centralized critic, while constraining actor policies to local observations at deployment. J. Tao, S. Xin, and F. Zhang are with the Robotics and Control Laboratory, School of Advanced Manufacturing and Robotics, and the State Key Laboratory of Turbulence and Complex Systems, Peking University, Beijing, 100871, China (email: [jiyuetao@pku.edu.cn](jiyuetao@pku.edu.cn); email: [shxin25@stu.pku.edu.cn](shxin25@stu.pku.edu.cn); email: [feitian@pku.edu.cn](feitian@pku.edu.cn)). T. Shen and D. Zhao are with the National Innovation Institute of Defense Technology, Beijing 100071, China (email: shents [bj@126.com](bj@126.com); email: [zhaodx2008@163.com](zhaodx2008@163.com)) . ∗ Send all correspondence to D. Zhao and F. Zhang. Yet even with centralized value guidance, actors trained end-to-end from local observations usually converge slowly and settle on suboptimal coordination strategies. A primary cause is partial observability [4], as the same local observation may be consistent with multiple global configurations, each requiring a different cooperative response. This ambiguity is further amplified under communication-free execution, where robots cannot exchange observations to reduce uncertainty. Such constraints are particularly prevalent in marine robotics, where underwater vehicles suffer from severe acoustic latency and packet loss [8], and surface vessels face inte

FAQs

What problem does CoDiMAD target in communication-free multi-robot coordination?

CoDiMAD targets slow convergence and suboptimal coordination caused by partial observability when robots cannot communicate, meaning local observations may correspond to multiple global configurations requiring different cooperative actions.

Why does standard deterministic distillation fail here?

Deterministic regression-based distillation collapses multi-modal conditional action distributions to their mean, which may not correspond to any valid coordinated action and can lead to hesitant or degenerate behavior.

How does CoDiMAD use diffusion to overcome mode averaging?

CoDiMAD trains a privileged oracle, builds an offline dataset of local observations and oracle actions, then distills the oracle into conditional denoising diffusion policies that sample from coherent coordination modes through the diffusion reverse process.

Jul 16, 2026

Upload by Valentina