💻 computer science

Distributed Model-Based Diffusion For Scalable Multi-Robot Trajectory Optimization

This paper proposes Distributed Model-Based Diffusion (DMBD), a scalable server-robot framework that decomposes the high-dimensional inference problem of multi-robot trajectory optimization into local conditional denoising processes, enabling efficient coordination in complex, non-convex environments with sub-second computation times.

Original authors: Haejoon Lee, Xinyi Wang, Taekyung Kim, Dimitra Panagou

Published 2026-09-16

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Haejoon Lee, Xinyi Wang, Taekyung Kim, Dimitra Panagou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the bustling world of robotics, getting a team of machines to move together without bumping into each other is a problem that has long frustrated engineers. Imagine a warehouse filled with dozens of autonomous carts, or a fleet of delivery drones navigating a crowded city skyline. Each machine needs to find a path to its destination, but the space is tight, the rules are complex, and the machines cannot simply guess their way through. If they rely on standard mathematical methods to calculate the perfect route, they often get stuck in local dead ends, unable to see a solution that requires a temporary detour or a coordinated pause. Furthermore, as the number of robots grows, the mathematical complexity explodes, often overwhelming the computers trying to solve the puzzle. This is where a newer approach, known as sampling-based optimization, steps in. Instead of trying to calculate a single perfect path through a rigid formula, these methods generate thousands of possible routes, test them against the rules, and gradually refine the best ones. It is a process of trial and error, but one that is surprisingly good at navigating the messy, unpredictable reality of the physical world.

A recent development in this field, called Model-Based Diffusion, took this idea a step further by borrowing a technique from generative artificial intelligence. Originally designed to create images by slowly removing noise from a blank canvas, this method was adapted to create robot paths. It starts with a chaotic, random set of movements and iteratively "denoises" them, smoothing out the trajectory until it becomes a viable, low-cost path that avoids obstacles. While this worked beautifully for a single robot, researchers at the University of Michigan discovered that applying it to a whole team was a different story. When they tried to use the method for multiple robots at once, the system required a central computer to hold the entire plan for every machine simultaneously. As the team size increased, the computer memory needed to process the joint plan grew so large that the system would crash, and the time required to find a solution became impractical. The central computer also needed to know the specific goals, physical limits, and internal rules of every single robot, which is often impossible in real-world scenarios where robots might be different from one another or keep their private objectives to themselves.

To solve this bottleneck, the researchers introduced a new framework called Distributed Model-Based Diffusion. Instead of forcing a single central brain to manage the entire team's plan, they broke the problem apart. In this new system, each robot acts as its own independent planner, but they all stay in touch through a simple server. The process works like a relay race of information. First, the server gathers the current best guess of where every robot is heading and broadcasts this summary to the entire team. Then, each robot uses this shared information to refine its own path. It asks itself, "Given where everyone else is going, what is the best way for me to move?" It generates its own set of potential paths, picks the best one based on its local goals and constraints, and sends its updated position back to the server. The server then aggregates these new positions and broadcasts them again for the next round of refinement. This cycle repeats, with the robots slowly converging on a coordinated set of paths without ever needing to know the full, complex details of their teammates' internal logic.

The researchers tested this approach in a series of rigorous simulations to see if it could handle the chaos of real-world coordination. They created scenarios ranging from a group of twenty robots swapping places in a crowded circle to a mixed fleet of different-sized vehicles navigating a multi-story building with elevators. In these tests, the new distributed method proved remarkably scalable. While older methods that tried to solve the problem all at once began to fail or run out of memory as the number of robots increased, the distributed approach maintained a steady speed. It could solve complex coordination tasks for twenty robots in less than a second, a feat that was impossible for the previous generation of algorithms. The system successfully managed tasks where robots had to temporarily move out of the way to let others pass, or where they had to navigate tight corners without colliding, all while only knowing their own specific goals and the general positions of their peers.

Perhaps the most significant finding was that the robots could achieve this high level of coordination without sharing their private objectives. In one simulation, a robot had to temporarily leave its own goal to allow another robot to park, a move that seemed counterintuitive for a machine focused only on its own destination. Yet, by relying on the shared trajectory estimates from the server, the robot naturally discovered this cooperative behavior. The system did not need to be explicitly programmed to be polite or strategic; the mathematical structure of the distributed denoising process naturally led the robots to find solutions that worked for the group as a whole. The researchers also provided a theoretical guarantee that the difference between this distributed approach and a perfect, centralized solution would shrink as the robots refined their paths, ensuring that the local decisions made by each robot would eventually align with the global optimum.

The results suggest that this method could be a turning point for large-scale robotic systems. By removing the need for a central computer to hold the entire map of the team's future, the approach opens the door to coordinating hundreds, or even thousands, of machines in dynamic environments. Whether it is a fleet of delivery drones avoiding each other in a busy sky or a swarm of rescue robots working together in a disaster zone, the ability to plan complex, collision-free paths in real-time without a single point of failure is a critical step forward. The work demonstrates that by letting each robot think for itself while staying loosely connected to the group, the team can solve problems that are too complex for any single mind to handle alone.

Technical Summary: Distributed Model-Based Diffusion for Scalable Multi-Robot Trajectory Optimization

1. Problem Statement

The paper addresses the challenge of collision-free trajectory optimization for multi-robot systems operating in highly non-convex, non-linear, and non-differentiable environments. While Model-Based Diffusion (MBD) has shown promise for single-robot trajectory generation by treating planning as a probabilistic inference problem, extending MBD to multi-robot systems presents two critical bottlenecks:

Curse of Dimensionality: In a centralized setting, the dimension of the joint trajectory space grows with the number of robots ( $N$ ). This leads to exponentially decreasing sampling efficiency and optimization performance as the team size increases.
Centralization Requirements: Standard MBD requires global access to all robots' dynamics, constraints, and objective functions to perform denoising. In practical, heterogeneous multi-robot systems, aggregating this information incurs substantial communication overhead and places a significant computational and memory burden on a single machine, limiting scalability.

Existing distributed sampling-based approaches often require agents to exchange multiple trajectory samples, resulting in high communication costs, or rely on gradient-based methods that assume differentiability and may get stuck in local minima.

2. Methodology: Distributed Model-Based Diffusion (DMBD)

The authors propose Distributed Model-Based Diffusion (DMBD), a server-robot framework that decomposes the global trajectory optimization problem into local conditional inference subproblems.

Core Mechanism

DMBD retains the forward (noising) process of standard MBD but fundamentally alters the reverse (denoising) process:

Decomposition: Instead of estimating a global score function over the full joint trajectory distribution, each robot $k$ independently estimates a local conditional score function.
Local Conditional Denoising: At each denoising step, robot $k$ updates its own control trajectory $Y_k$ using only its local objective $J_k$ , local dynamics $f_k$ , and local constraints $C_k$ .
Conditioning: The local update is conditioned on the current state trajectory estimates of all other robots ( $\tau_{-k}$ ). These estimates are aggregated and broadcast by a central server.
Communication Efficiency: Unlike other distributed methods, each robot transmits only a single trajectory per denoising step to the server, rather than multiple samples. The server performs no optimization; it only aggregates and rebroadcasts trajectories.

Algorithmic Flow (Algorithm 1)

Initialization: Each robot samples an initial noisy trajectory $Y^{(M)}_k$ and sends its state trajectory $\tau^{(M)}_k$ to the server.
Aggregation: The server aggregates all trajectories into a global state $\tau^{(M)}$ and broadcasts it back.
Iterative Denoising ( $i = M \dots 1$ ):
- Each robot $k$ receives the trajectories of others ( $\tau^{(i)}_{-k}$ ).
- Robot $k$ samples $S_k$ candidate trajectories from a Gaussian distribution centered at its current estimate.
- It computes a weighted average $\hat{Y}^{(i)}_k$ based on the local conditional likelihood $p^k_0(Y_k | \tau_{-k})$ .
- The robot updates its trajectory using the estimated local score function (Eq. 17) and sends the new state trajectory back to the server.
Termination: The process repeats until $i=1$ , yielding optimized trajectories $\{Y^{(0)}_k\}$ .

Theoretical Analysis

The paper establishes a theoretical bound on the discrepancy between the centralized global score function and the distributed local conditional score function (Proposition 1).

The discrepancy is bounded by the Lipschitz constant of the local denoiser and the distance between the current estimates of other robots' trajectories and their fully denoised counterparts.
The bound implies that as the denoising process converges (estimates approach the true solution) or if constraints are independent, the distributed approach converges to the centralized solution.
Crucially, the method is zeroth-order and does not assume differentiability of objectives or constraints.

3. Key Contributions

Proposal of DMBD: A server-coordinated distributed algorithm that decomposes multi-robot trajectory optimization into local conditional diffusion processes, enabling each robot to denoise its trajectory using only local information conditioned on others.
Theoretical Bounds: A derivation of the relationship between MBD and DMBD, providing a theoretical bound on the discrepancy between global and local conditional score functions.
Empirical Validation: Extensive simulations demonstrating superior scalability and performance across diverse scenarios compared to existing baselines.

4. Experimental Results

The authors evaluated DMBD against four sampling-based optimization (SBO) baselines: Cross-Entropy Method (CEM), Model Predictive Path Integral (MPPI), standard MBD, and D4orm.

Scenarios

Goal Swapping: $N$ robots (circular and rectangular) swapping positions in a circle.
Multi-Floor Coverage: Heterogeneous teams navigating a two-story building with elevator constraints.
Parking: Two rectangular robots coordinating to park where one must temporarily vacate a goal to let the other in.
Rush Hour: A four-robot puzzle requiring strategic coordination to clear a path for one robot.

Performance Metrics

Scalability (Q1): DMBD maintains near-constant planning time as the number of robots ( $N$ ) increases. In contrast, baselines suffer from rapidly increasing planning times and decreasing success rates. Notably, all baselines failed (out-of-memory errors) at $N=20$ for goal-swapping tasks, whereas DMBD succeeded.
Success Rate: DMBD achieved the highest success rates in complex tasks (rectangular robots, multi-floor coverage), often exceeding 90%.
Coordination without Global Knowledge (Q2): In the "Parking" and "Rush Hour" scenarios, robots achieved high success rates (>90%) without access to global objectives or dynamics. This demonstrates that local conditional denoising can emergently produce globally coordinated behavior.
Computation Time: DMBD solved challenging coordination tasks with sub-second computation times (e.g., 0.39s for Rush Hour, 0.52s for Parking).

5. Significance and Claims

The paper claims that DMBD successfully addresses the scalability limitations of Model-Based Diffusion in multi-robot systems. By decomposing the inference problem, DMBD avoids the curse of dimensionality associated with joint trajectory spaces and eliminates the need for a central processor to have full knowledge of the system.

The significance lies in enabling scalable, learning-free, and differentiable-free trajectory optimization for large teams of heterogeneous robots. The method retains the multi-modal trajectory generation capabilities of diffusion models while operating efficiently in distributed settings where global information is unavailable or communication is constrained. The results suggest that DMBD is a viable solution for real-world coordination tasks involving complex constraints and large robot teams.