Rapid and Safe Trajectory Planning over Diverse Scenes through Diffusion Composition

Anonymous Authors
Anonymous Institution

Teaser

Diffusion Composition Enables Efficient, Safe Planning with Practical Real-world Performance. An individual diffusion model cannot ensure safe trajectory planning in out-of-distribution scenarios, whereas composing multiple diffusion models can achieve safety during generalization. Dashed boxes indicate obstacles that do not exist during training. Validation on the F1TENTH platform shows that trajectories planned by the composed diffusion model offer excellent safety while maintaining computational efficiency, demonstrating effectiveness for practical real-world applications.

Abstract

Safe trajectory planning in complex environments must balance stringent collision avoidance with real-time efficiency, which is a long-standing challenge in robotics. In this work, we present a diffusion-based trajectory planning framework that is both rapid and safe. First, we introduce a scene-agnostic, MPC-based data generation pipeline that efficiently produces large volumes of kinematically feasible trajectories. Building on this dataset, our integrated diffusion planner maps raw onboard sensor inputs directly to kinematically feasible trajectories, enabling efficient inference while maintaining strong collision avoidance. To generalize to diverse, previously unseen scenarios, we compose diffusion models at test time, enabling safe behavior without additional training. We further propose a lightweight, rule-based safety filter that, from the candidate set, selects the trajectory meeting safety and kinematic-feasibility requirements. Across seen and unseen settings, the proposed method delivers real-time-capable inference with high safety and stability. Experiments on an F1TENTH vehicle demonstrate practicality on real hardware.

Real-World Validation of Individual Static Model

replay

replay
replay

replay

replay
replay
This paper introduces a fast trajectory‐synthesis framework that seamlessly fuses real‐time perception with diffusion‐based generative models. By distilling rich sensor inputs into compact vehicle‐state embeddings, our approach dramatically accelerates the iterative denoising process while fully preserving the diffusion model's capacity to produce precise, high‐fidelity trajectories.

Real-World Validation of Composition for UNSEEN Scenes

Static Diffusion Model replay

Dynamic Diffusion Model replay
Composed Result replay
This model composition approach ensures collision-free trajectory planning in unseen scenes for safety. The proposed method can make safe test-time decisions to generate safe behaviors, such as decelerating to avoid obstacles, thus allowing collision-free trajecoty planning for diverse scenes without retraining.

Composed Scene 1 - view 1:

Initial Pose: [0.86, 0.03] replay

Initial Pose: [2.73, 0.32] replay
Initial Pose: [3.03, -0.07] replay

Composed Scene 1 - view 2:

Initial Pose: [0.86, 0.03] replay

Initial Pose: [2.73, 0.32] replay
Initial Pose: [3.03, -0.07] replay

Composed Scene 2 - view 1:

Initial Pose: [0.82, 0.07] replay

Initial Pose: [2.76, 0.47] replay
Initial Pose: [2.79, -0.11] replay

Composed Scene 2 - view 2:

Initial Pose: [0.82, 0.07] replay

Initial Pose: [2.76, 0.47] replay
Initial Pose: [2.79, -0.11] replay

The Performance of Model Composition in Simulation

Weights of composition - [Dynamic Model, Static Model]:  
Static Model

Dynamic Model

Composed Result
Effectiveness of Models Composition and Sensitivity of Compositional Weights. The composed model can safely generalize to unseen scenes when appropriate compositional weights are selected, achieving collision-free trajectory planning.