DiRecT Enhances Safe Diffusion-Based Planning with Receding-Horizon Denoising

Diffusion models have emerged as powerful tools for planning and control, capable of learning complex, multimodal distributions over actions and trajectories. However, a significant hurdle to their deployment in safety-critical applications is the reliable enforcement of safety constraints during inference. Current methods often apply these constraints to each intermediate denoising step, which can inadvertently overconstrain the sampling process and degrade the quality of the generated samples. To overcome this limitation, researchers introduce DiRecT, which stands for Diffusion-based planning via Receding-horizon denoising with Terminal constraints. This training-free algorithm is designed to enforce constraints exclusively on the final, clean trajectory, thereby avoiding unnecessary restrictions on the noisy intermediate denoising dynamics. Inspired by model predictive control, DiRecT derives a principled receding-horizon surrogate for the otherwise intractable constrained stochastic optimal control (SOC) formulation. This results in an efficient algorithm that distinctly separates stochastic denoising from constraint satisfaction, progressively guiding samples towards feasible final trajectories without distorting the learned diffusion dynamics. The flexibility of DiRecT allows it to integrate various optimizers, incorporate environmental dynamics priors, and optimize additional soft rewards. Extensive experiments on safe planning benchmarks demonstrate that DiRecT substantially improves both deployment safety and task performance compared to existing diffusion-based planning baselines.