CoDiMAD: Diffusion-Based Privileged Distillation for Communication-Free Multi-Robot Coordination
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "CoDiMAD: Diffusion-Based Privileged Distillation for Communication-Free Multi-Robot Coordination".
Dev: Here's the gist of the paper. Decentralized multi-robot coordination under partial observability remains challenging, especially in communication-free settings where agents must act solely from local sensor observations.
Rosa: Privileged policy distillation provides a promising approach by transferring knowledge from a globally informed oracle to sensor-constrained students.
Dev: CoDiMAD proposes a three-stage framework that trains an oracle with MAPPO, constructs an offline dataset of local observation-oracle-action pairs, and distills the oracle into decentralized students parameterized as conditional denoising diffusion probabilistic models.
Rosa: 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.
Title and authors: Rosa: This paper is titled "CoDiMAD: Diffusion-Based Privileged Distillation for Communication-Free Multi-Robot Coordination". The authors are Jiyue Tao, Shunheng Xin, Tongsheng Shen, Dexin Zhao, and Feitian Zhang.
Dev: They are coming from different corners of this stuff—robotics, control engineering, autonomy research—which is good because it means they’re looking at the problem from all angles.
Taro: I see a lot of complexity in their setup with that dual-stream input for the oracle policy: local maps and global maps. It sounds like a lot of data to process before you get to the action output.
Rosa: Yeah, they use parallel CNN branches for both streams, then aggregate those temporal features using a recurrent unit called a GRU before sending it through an MLP head to decide what action to take.
Dev: That centralized MAPPO critic during training is key because it gives the oracle policy privileged value estimates, which is how they get that high-quality initial knowledge base.
The paper's summary: Rosa: So, the summary of this paper explains the three main stages they follow. First, training an oracle policy with MAPPO using both local and global maps. Then, collecting an offline dataset of local observation and oracle action pairs from all agents.
Dev: And finally, distilling that knowledge into a decentralized student policy parameterized as conditional denoising diffusion probabilistic models. That’s the whole pipeline they propose for communication-free coordination.
Taro: The part I find really interesting is how they formally show that partial observability causes these multi-modal action distributions where standard regression distillation fails because it just finds the mean, which isn't a real action.
Rosa: Right, so deterministic methods collapse those modes into one point, and CoDiMAD uses the diffusion reverse process to sample actions from those distinct modes instead of averaging them out.
Dev: That means they aren't just getting a hesitant action; they are sampling from coherent coordination behaviors that exist in the system.
The paper's improvements: Taro: The theoretical analysis shows that if the different modes of actions are well separated, then the MSE-optimal deterministic student policy converges to a point between those modes, which means it lands in an invalid or hesitant region.
Rosa: That’s exactly why they say diffusion modeling recovers these individual modes by training the model to estimate the conditional score of that noisy action distribution. It learns how to sample from all of them.
Dev: And they mention integrating Denoising Diffusion Implicit Models, or DDIM, for accelerated sampling. This is important because diffusion models can be slow during deployment, so this makes it feasible for real-time onboard use in communication-free settings.
Rosa: So the improvements are really about moving from a single best guess to a distribution of good guesses that actually capture the system's complex behavior.
Taro: They validate this on three tasks—Cooperative Coverage, Pursuit-Evasion, and Box Pushing—showing consistent improvements over direct local MARL and deterministic distillation baselines.
Conclusion: Dev: So to wrap up, the main point of "CoDiMAD: Diffusion-Based Privileged Distillation for Communication-Free Multi-Robot Coordination" is that using diffusion modeling lets decentralized students sample from coherent coordination modes instead of averaging across them.
Rosa: It successfully recovers those distinct modes, and experimentally, it consistently beats deterministic distillation baselines on three different cooperative tasks. For example, on Cooperative Coverage, CoDiMAD hit a ninety-five point seven percent coverage rate compared to the oracle’s ninety-seven point six percent.
Taro: I just want to add that the paper notes their limitation in this approach is that it still relies on having an offline dataset of local observation and oracle action pairs to start with.
Dev: That makes sense, you need that training data before you can distill anything, right? But they also mention incorporating a recurrent unit like a GRU into the student encoder to help it handle the temporal history better for sequential tasks.
Rosa: So we’ve seen how this privileged distillation framework works and how diffusion sampling lets us capture multi-modal coordination without relying on direct communication between robots. We’ll take a quick break and come back to look at some other interesting papers.
Peking University · National Innovation Institute of Defense Technology
cs.RO
Submitted: 2026-07-10
Updated: 2026-11-05
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 92/100
The gist: Decentralized multi-robot coordination under partial observability remains challenging, especially in communication-free settings where agents must act solely from local sensor observations.
Key concepts
- Dec-POMDP
- Decentralized Partially Observable Markov Decision Process models multi-robot coordination where agents act based only on their local sensors. The challenge is that the same local observation can mean different global situations, leading to uncertain and potentially incorrect cooperative actions.
- Privileged Policy Distillation
- This technique transfers knowledge from a globally informed 'oracle' policy, which has access to all information, to sensor-constrained 'student' policies. The goal is for the student to learn complex coordination behaviors without needing constant communication with the oracle during operation.
- Conditional Denoising Diffusion Models
- The student policy is parameterized as a diffusion model trained to predict noise in a process that corrupts the clean action. This allows it to approximate the complex conditional distribution of actions given local observations, enabling it to sample from multiple valid coordination modes instead of just averaging them.
Terminology
Summary
Decentralized multi-robot coordination under partial observability remains challenging, especially in communication-free settings where agents must act solely from local sensor observations. p. 0 Privileged policy distillation provides a promising approach by transferring knowledge from a globally informed oracle to sensor-constrained students. p. 0 p. 0 CoDiMAD proposes a three-stage framework that trains an oracle with MAPPO, constructs an offline dataset of local observation-oracle-action pairs, and distills the oracle into decentralized students parameterized as conditional denoising diffusion probabilistic models. p. 0 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. p. 0 p. 0
The Core Problem and Proposed Solution
Multi-robot coordination under partial observability is modeled as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP). The core difficulty arises because the same local observation may correspond to multiple global configurations requiring qualitatively different cooperative actions, making the conditional action distribution inherently multi-modal. p. 0 Standard deterministic distillation collapses these modes to their mean, often yielding invalid or hesitant actions. p. 0 p. 0 CoDiMAD addresses this by modeling the conditional action distribution rather than a point estimate. p. 0
The Three-Stage Framework
CoDiMAD is structured as a three-stage oraclestudent distillation framework.
-
Stage 1 involves training a shared oracle policy πO via MAPPO with access to both local and privileged global observations. p. 2
-
Stage 2 focuses on collecting an offline dataset D pairing local-observation with oracle actions from all agents instances and timesteps. p. 2
-
Stage 3 distills the oracle's behavior into a shared diffusion-based student policy πθ conditioned solely on local observation history.
Theoretical Advantages of Diffusion Modeling
The theoretical analysis formalizes the issue by showing that partial observability induces multi-modal action distributions under which regression-based distillation suffers from mode averaging. Proposition 1 demonstrates that if class means are well separated, the conditional oracle-action distribution pD(a oi) is multi-modal. p. 4 Corollary 1 proves that the MSE-optimal deterministic student converges to a point between modes, leading to low density in high-probability regions. p. 0 Diffusion sampling recovers these individual modes by training the model to estimate the conditional score of the noised action distribution qk(a(k) c). p. 4
Experimental Validation and Results
Experiments on three cooperative tasks—Cooperative Coverage, Pursuit-Evasion, and Box Pushing—show that CoDiMAD consistently outperforms direct local MARL and deterministic distillation baselines. For instance, on Cooperative Coverage, CoDiMAD achieves a 95. p. 57% coverage rate compared with the oracle’s 97. p. 5 p. 56%. Furthermore, CoDiMAD consistently reduces collisions from 10. p. 537 to 0.55 per episode on Coverage, an approximately 19× reduction compared to BC-RNN. p. 5 The results validate the ability of diffusion-based distillation to transfer a substantial portion of the globally-informed cooperative behavior to a sensor-constrained decentralized student. p. 5
Conclusion and Future Work
CoDiMAD successfully recovers distinct modes through diffusion sampling, demonstrating that local observations can correspond to multi-modal oracle action distributions. The framework consistently outperforms deterministic distillation baselines across all tested tasks. p. 7 Future work may involve incorporating online fine-tuning with diffusion-compatible RL objectives to improve robustness under distribution shifts. p. 7 Scalability beyond N = 3 agents and deployment on physical marine robots are also suggested directions for future research. p. 7
REFERENCES
CoDiMAD: Diffusion-Based Privileged Distillation for Communication-Free Multi-Robot Coordination
arXiv:2607.09587v1 [cs. p. 0RO] 10 Jul 2026
--- Page 1 ---
CoDiMAD: Diffusion-Based Privileged Distillation for Communication-Free Multi-Robot Coordination
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. p. 0 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. p. 0 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. p. 0
--- Page 2 ---
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. p. 0 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. p. 0 Yet even with centralized value guidance, actors trained end-to-end from local observations usually converge slowly and settle on suboptimal coordination strategies. p. 0 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. p. 0 This ambiguity is further amplified under communication-free execution, where robots cannot exchange observations to reduce uncertainty. p. 0 Such constraints are particularly prevalent in marine robotics, where underwater vehicles suffer from severe acoustic latency and packet loss [8], and surface vessels face intermittent, range-limited radio-frequency links [9]. p. 0 A principled way to mitigate partial observability is privileged policy distillation [10]. p. 0 Instead of learning coordination directly under limited observations, one first trains an oracle policy with full state access, where the learning problem is substantially easier, and then transfers its behavior to sensor-constrained students via behavioral cloning (BC). p. 0 This paradigm has achieved notable success in single-robot settings, including vision-based driving distilled from a map-equipped oracle [11] and proprioceptive locomotion distilled from a terrain-aware teacher [12]. p. 0 Inspired by these successes, we extend privileged distillation to multi-robot coordination. p. 0 However, this extension faces a fundamental challenge that is largely absent in single-robot tasks. p. 0 In cooperative multi-agent systems, the oracle action for an individual robot may depend on the joint configuration of all agents and environment states beyond that robot’s sensing range. p. 0 Consequently, the same local observation can correspond to multiple distinct oracle actions, inducing an inherently multi-modal conditional action distribution (Proposition 1). p. 0 Standard deterministic policies cannot represent such distributions. p. 0 Under common regression-based BC objectives such as mean squared error (MSE), they converge to the conditional mean, which may lie between distinct modes and correspond to no valid cooperative action (Corollary 1). p. 0
--- Page 3 ---
B. Observation Spaces and Information Asymmetry
In this work, we instantiate the above formulation for a team of homogeneous robots. To represent spatial and task-relevant information in a unified format, all local and global observations are structured as multi-channel bird’s-eye-view (BEV) grid maps with a spatial resolution of 32×32, as summarized in Table I. p. 1 The privileged oracle policy processes a dual-stream input (o loc i, o glo i). p. 1 The local stream o loc i ∈ R 2×32×32 consists of an egocentric LiDAR occupancy map and a velocity-field map. p. 1 The privileged global stream o glo i ∈ R 3×32×32 contains three channels that encode arena-scale global information, including the absolute positions of the self-agent, its teammates, and the full task-state map (e. p. 2g., global target locations). p. 2 In contrast, the student policy operates under realistic sensing limitations and observes only a singlestream tensor o i ∈ R 4×32×32, where the three global channels are replaced by two locally-restricted channels—a cropped task-state slice and teammate positions clipped to the agent’s limited sensing radius, which yields a 4-channel input combined with the two local channels from o loc i. p. 2 This information asymmetry is the core source of multi-modal uncertainty. p. 2
--- Page 4 ---
IV. METHODOLOGY
A. Stage 1: Oracle Policy Training via MAPPO
We train the oracle policy πO following the CTDE paradigm using MAPPO. In our homogeneous robot domains, the oracle actor is shared across agents and is evaluated independently for each agent. p. 2 The actor receives the dual-stream observation (o loc i, o glo i) comprising both local and global map inputs. p. 2 During deployment, only the actor is retained for offline data collection. p. 2
--- Page 5 ---
B. Stage 2: Offline Trajectory Collection
Once the oracle policy converges, it is frozen and executed deterministically by taking the mean of its Gaussian action distribution. For each rollout step, we record (o t i, a t i, dt i) for all agents i and timesteps t. p. 2 Importantly, o t i excludes the oracle’s privileged global stream and contains only the information available to the student at deployment. p. 2
--- Page 6 ---
C. Stage 3: Diffusion-Based Behavioral Cloning
-
Architecture: The student policy πθ is a conditional denoising diffusion model comprising an observation encoder and a noise prediction network ϵˆθ.
-
Training Objective: Let a(0) denote the normalized oracle action. For notational simplicity, we omit the agent and time indices in the following diffusion equations. p. 3 The forward diffusion process corrupts the clean action a(0) by systematically adding Gaussian noise over T = 200 timesteps according to a cosine noise schedule. p. 3 To train the student policy πθ to mimic the oracle’s expertise under partial observability, we optimize the noise prediction network ϵˆθ following the DDPM framework [15]. p. 3
--- Page 7 ---
D. Theoretical Analysis
Proposition 1 (Multi-modality from partial observability)
Corollary 1 (Mode averaging under deterministic distillation).
--- Page 8 ---
III. PROBLEM FORMULATION
A. Dec-POMDP Formulation
We model multi-robot coordination as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP).
--- Page 9 ---
V. EXPERIMENTS
A. Evaluation Environments
• Cooperative Coverage: Agents explore the arena to maximize coverage of the occupancy grid.
• Pursuit-Evasion: Pursuers cooperate to capture a single evader controlled by an Artificial Potential Field (APF) policy. p. 4
• Box Pushing: Agents cooperatively push a heavy box toward a randomly placed goal zone. p. 4
--- Page 10 ---
TABLE II
PERFORMANCE COMPARISON ACROSS THREE COOPERATIVE TASKS.
--- Page 11 ---
CoDiMAD vs. MAPPO-Oracle
CoDiMAD closely approaches the privileged oracle on two of the three tasks. On Coverage, CoDiMAD achieves 95. p. 5 p. 57% coverage rate compared with the oracle’s 97.6%. On Pursuit-Evasion, CoDiMAD reaches a 90. p. 56% capture rate compared to the oracle’s 99. p. 51%, with similar collision frequency. p. 5 On Box Pushing, CoDiMAD achieves 72. p. 52% success compared to the oracle’s 98. p. 52%, which is the largest gap among the three tasks. p. 5
--- Page 12 ---
Diffusion vs. regression (CoDiMAD vs. p. 0 BC-RNN
CoDiMAD consistently outperforms BC-RNN, supporting the motivation developed in Section IV-D. On Coverage, CoDiMAD achieves 95.7% coverage rate compared with BC-RNN’s 80. p. 59%, corresponding to a 14. p. 58 percentage-point improvement. p. 5 At the same time, it reduces collisions from 10. p. 537 to 0.55 per episode, an approximately 19× reduction. p. 5 On Pursuit-Evasion, the two methods achieve comparable capture rates (90. p. 56% vs. 90.3%), but CoDiMAD reduces collisions from 3. p. 541 to 0.80, a 4.3× improvement. p. 5 On Box Pushing, the advantage of CoDiMAD becomes more pronounced: it achieves 72. p. 52% success rate compared to BC-RNN’s 6. p. 55%.
--- Page 13 ---
Temporal history (CoDiMAD vs. CoDiMAD w/o RNN
Removing the GRU leads to a moderate but consistent performance drop across all three tasks. The effect is most visible on Pursuit-Evasion, where the capture rate decreases from 90. p. 66% to 80.4% and collisions increase from 0. p. 680 to 1.41 per episode. p. 5
--- Page 14 ---
D. Analysis of Multi-Modal Behavior
Action Distribution Visualization Visualization
Agent 1 Agent 2 Agent 3 (a) BC-RNN (Coverage) (b) CoDiMAD (Coverage) (c) BC-RNN (Pursuit) (d) CoDiMAD (Pursuit).
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Diffusion-Based Multi-Agent Distillation for Communication-Free Coordination: The improved system can train a decentralized student policy that samples from coherent coordination modes rather than averaging across them, addressing the
mode-averaging failure of deterministic distillation
by approximating the conditional oracle action distribution through the diffusion reverse process. -
Real-Time Onboard Deployment via DDIM Integration: By integrating Denoising Diffusion Implicit Models (DDIM), the system can achieve accelerated, few-step sampling of actions, enabling real-time onboard deployment suitable for communication-free environments where high inference latency is typically prohibitive.
-
Enhanced Robustness Against Partial Observability: The system can recover distinct coordination modes under partial observability by modeling the conditional action distribution rather than a point estimate, allowing agents to sample from
coherent coordination modes
instead of collapsing to an invalid or hesitant action. -
Superior Task Performance via Privileged Distillation: The system can consistently outperform direct local MARL and deterministic distillation baselines across cooperative tasks, such as achieving 95.7% coverage rate compared to the MAPPO-Oracle's 97.6% in the Coverage task, while maintaining a comparably low collision count.
-
Improved Coordination Precision in Sensitive Tasks: For tasks requiring precise coordination like Box Pushing, the system can achieve a significantly higher success rate (72.2%) compared to deterministic regression methods (6.5%), demonstrating its ability to capture
tight force-coordination requirements.
-
Contextual Awareness through Temporal History Encoding: Incorporating a GRU in the student encoder allows the system to maintain a recurrent hidden state summarizing observation history, which is shown to provide
useful contextual information for disambiguating single-frame observations
and improves performance on sequential coordination tasks.
Sources
- KD-MARL: Resource-Aware Knowledge Distillation in Multi-Agent Reinforcement Learning
- Interactive Distillation for Cooperative Multi-Agent Reinforcement Learning
- MIMIC-D: Multi-modal Imitation for MultI-agent Coordination with Decentralized Diffusion Policies
- IDQL: Implicit Q-Learning as an Actor-Critic Method with Diffusion Policies
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation