Denoising Multi-Robot Trajectories

What it claims

D4ORM is a learning-free, GPU-parallel trajectory optimizer for multi-robot teams. It is coupled, offline, full-horizon and centralized. Its main limitation is a scalability bottleneck: as robot count grows, the gradient signal for the joint team trajectory gets diluted. This paper analyzes that bottleneck and wraps D4ORM in three variants (Table I):

• D4ORM-D for scalability
• D4ORM-C2F for real-time receding-horizon use
• D-D4ORM for onboard runs without GPUs

The paper is accepted to IEEE Transactions on Robotics (September 2026).

How it works

• Base D4ORM. It builds on model-based diffusion (MBD). It does not sample trajectories from scratch. Instead it repeatedly optimizes deformations, which are incremental updates to the current joint control trajectory for the whole team. Each iteration runs a full diffusion-denoising pass.
• Each denoising step uses an on-the-fly Monte Carlo gradient estimate of an interpretable objective. No analytical gradients or trained model are needed.
• Robot dynamics can be arbitrary and mixed, e.g. double-integrator, holonomic and differential-drive in one team.
• It is anytime: solutions improve the longer it runs.
• D4ORM-D. A semi-decoupled method that resolves conflicts iteratively by running D4ORM on subsets of the team.
• D4ORM-C2F. D4ORM-D produces coarse reference trajectories offline. Online, D4ORM generates fine trajectories that track those references, inside a receding-horizon loop with feedback control.
• D-D4ORM. Planning is split across robots with inter-robot communication. This allows onboard execution, including asynchronous target updates for "lifelong" navigation.

flowchart LR
  B[D4ORM base] --> D[D4ORM-D subsets]
  D -->|coarse refs, offline| C[D4ORM-C2F online]
  B -->|fine tracking| C
  B --> X[D-D4ORM onboard, no GPU]

Evidence

The fetched text stops in the related-work section, so no results tables are available. The only quantitative figure is in Fig. 1e: D4ORM-D plans for 100 holonomic 2D robots in about 30 s. That run optimizes 2 DoF × 250 horizon steps × 100 robots of variables. Hardware and baseline timing are not given here.

The claim of beating MPPI and a learned diffusion method is stated only at abstract level.

For context, the authors note that prior sampling-based multi-robot methods use robot-wise decoupled, limited-horizon representations and generally handle about a dozen robots.

How to use it

Code and video are at https://github.com/proroklab/d4orm. The page states no licence, install steps or API.

Use the variant that matches your constraint:

Limits

• Scaling. Vanilla D4ORM degrades with team size because of gradient dilution. The authors characterize this as a limitation and say it motivates the variants (Section IV-B).
• Coordination loss. The decoupled and receding-horizon variants give up some of the full joint search. The paper's own framing is that decoupling restricts the solution space and can prevent tightly coordinated long-term behavior.
• Missing results. The fetched text omits all results, including success rates, solve times against MPPI and MBD, and what "zero-shot" quadrotor deployment involved. Read Section VIII before relying on the performance claims.