Research map / Image generation / Fast diffusion sampling

DiRecT: Safe Diffusion-Based Planning via Receding-Horizon Denoising

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Guidance for conditional diffusion · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DiRecT, a training-free constrained sampler for diffusion planners that enforces hard constraints only on the final clean trajectory via a receding-horizon SOC surrogate, improving safety and task success on Maze2D, D3IL, MRMP and PushT.

Problem

Diffusion-based planners must produce trajectories that satisfy hard safety constraints at inference time, but existing methods enforce feasibility on noisy intermediate denoising iterates, which overconstrains the sampler and degrades sample quality.

Approach

Formulate constrained diffusion sampling as a terminally constrained stochastic optimal control problem and derive a one-step receding-horizon surrogate: at each denoising step, solve a constrained optimization on the Tweedie clean-sample prediction and map the correction back to the noisy latent via a scaled difference of predicted clean samples.

Key results

• On Maze2D Broad, DiRecT reaches safety rate 0.970 versus 0.220 for the best baseline (Projected Diffusion).
• On D3IL Avoiding, DiRecT achieves SR 1.000 and task success 1.000 versus 0.880 SR for Gradient Guidance.
• On Maze2D Narrow, DiRecT attains SR 0.940 with near-zero violations, while Projected Diffusion reaches 0.780 SR with a lower score.

Surprising findings

• Projection-based baselines such as Augmented Lagrangian and SafeDiffuser-RoS show low safety rates despite their constraint-enforcement design, and Diffuser unguided has SR 0.000 on Maze2D.

Limitations

• Inference-time overhead from per-step constrained optimization; sampling time on Maze2D is 17.7s vs 0.3s for Diffuser.
• Relies on approximations (one-step MPC surrogate, one-step fixed-point inversion of Tweedie) without full theoretical guarantees beyond terminal feasibility.
• Requires a domain-specific or off-the-shelf optimizer for the projection subproblem.

Experimental setup

Datasets: D4RL Maze2D-large D3IL avoiding demonstrations MMD multi-robot motion planning benchmark PushT

Benchmarks: Maze2D Broad Maze2D Narrow D3IL Avoiding MMD multi-robot motion planning (Empty, Highways, Conveyor, Drop-Region) PushT coupled generation

Metrics: safety rate violations D4RL score sampling time task success success rate constraint safety data adherence DTW DFD

Baselines: Diffuser Gradient Guidance Projected Diffusion Augmented Lagrangian SafeDiffuser-RoS MMD-CBS PCD-LB PCD-SHD Diffusion Policy CD-DPP CD-LB PCD-DPP

Builds on: Diffuser Projected Diffusion SafeDiffuser model predictive control Tweedie's formula

Closest papers by meaning

• SafeDiffuser: Safe Planning with Diffusion Probabilistic Models (ICLR 2025 · similarity 0.90)
• Constrained Diffusers for Safe Planning and Control (NeurIPS 2025 · similarity 0.90)
• Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 2024 · similarity 0.90)
• Prior-Guided Diffusion Planning for Offline Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
• Safe and Stable Control via Lyapunov-Guided Diffusion Models (NeurIPS 2025 · similarity 0.88)
• DiffuserLite: Towards Real-time Diffusion Planning (NeurIPS 2024 · similarity 0.88)