Research map / Image generation / Fast diffusion sampling
SafeDiffuser: Safe Planning with Diffusion Probabilistic Models
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.
Introduces SafeDiffuser, which enforces safety specifications on diffusion-based planners by embedding finite-time diffusion invariance via control barrier functions and a per-step QP projection.
Problem
Diffusion-based trajectory planners (diffusers) provide no safety guarantees, which blocks their use in safety-critical planning.
Approach
Treat the denoising diffusion time as a control dynamic and apply control barrier functions along diffusion time so that specifications hold at the final (noise-free) diffusion step; three variants (robust-safe, relaxed-safe, time-varying-safe) are enforced with a minimum-deviation QP at every denoising step.
Key results
- On maze planning, RoS, ReS and TVS diffusers reach near-zero violation (about -2.4e-7 on simple and complex specs) versus -0.983 and -0.894 for the vanilla Diffuser, with TVS scoring 0.806 vs 1.016 for Diffuser.
- On Walker2D and Hopper, the RoS-diffuser achieves zero collision violations on simple specs, while the vanilla Diffuser has -9.375 and -2.180 violations.
- On Pybullet manipulation, RoS-diffuser reward is 0.925 versus 0.650 for the Diffuser baseline, with satisfied specs (0.072 S-spec, 0.069 C-spec).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add robust-safe CBF QP enforcement inside diffusion steps | Diffuser on maze planning · Maze (Table 1) | Score: 1.016 to 0.770 (lower score for RoS); mixed comparison | Helped |
| Add robust-safe CBF QP enforcement inside diffusion steps | Diffuser on maze planning · Maze safety satisfaction | S-spec and C-spec satisfaction (≥0): S-spec -0.983 to -2.384e-7; C-spec -0.894 to -5.96e-7 | Helped |
| Time-varying-safe CBF enforcement (diffusion-time-varying constraint) | Diffuser on maze planning · Maze (Table 1) | Score: 0.806 vs 1.016 baseline; not higher than baseline | Helped |
| Relaxed-safe CBF with relaxation variables and extra diffusion steps | Diffuser on maze planning · Maze (Table 1) | Score: 0.762 vs 0.770 for RoS | No clear effect |
| Truncation to enforce constraints | Diffuser on maze planning · Maze (Table 1) | S-spec and C-spec satisfaction: S-spec satisfied (-1.192e-7) but C-spec not (-0.759); score 0.754 | Mixed |
| Classifier guidance for safety | Diffuser on maze planning · Maze (Table 1) | Score: 0.502 vs 1.016 baseline | Hurt |
| Robust-safe CBF enforcement | Diffuser on Walker2D locomotion · Walker2D (Table 2) | Score: 0.346 to 0.312 | Hurt |
| Robust-safe CBF enforcement | Diffuser on Hopper locomotion · Hopper (Table 2) | Score: 0.455 to 0.430 | Hurt |
| Robust-safe CBF enforcement | Diffuser on manipulation · Pybullet manipulation (Table 3) | Reward: 0.650 to 0.925 | Helped |
| Robust-safe CBF enforcement | Diffuser on manipulation · Pybullet manipulation C-spec | C-spec satisfaction (≥0): -0.065 to 0.069 | Helped |
| Truncation for specifications | Diffuser on Walker2D and Hopper · Locomotion (Table 2) | Score: Walker2D 0.346 to 0.286; Hopper 0.455 to 0.436 | Hurt |
Surprising findings
- On manipulation, the safety-enforcing RoS-diffuser improves reward over the unconstrained Diffuser (0.925 vs 0.650).
- On maze planning, the maze-score of the maximally-safe TVS-diffuser (0.806) is slightly below the unconstrained Diffuser (1.016) only in terms of score ordering; RoS-diffuser scores 0.770, so safety costs some performance in some settings.
Limitations
- Specifications must be continuously differentiable and known in advance.
- Satisfaction scores are not strictly non-negative due to numerical errors and inter-sampling effects.
- Robust-safe variant can leave trajectory points stuck in local traps at the boundary.
Experimental setup
Datasets: D4RL Maze2D D4RL Walker2D-Medium-Expert D4RL Hopper-Medium-Expert Pybullet manipulation trajectories
Benchmarks: Maze2D Walker2D Hopper Pybullet manipulation
Metrics: specification satisfaction (S-spec, C-spec) task score reward computation time per diffusion step
Baselines: Diffuser Truncate Classifier guidance Classifier guidance-epsilon
Builds on: Diffuser control barrier functions high-order CBF denoising diffusion probabilistic models
Closest papers by meaning
- DiRecT: Safe Diffusion-Based Planning via Receding-Horizon Denoising (NeurIPS 2026 · similarity 0.90)
- Constrained Diffusers for Safe Planning and Control (NeurIPS 2025 · similarity 0.88)
- Model-based Diffusion for Trajectory Optimization (NeurIPS 2024 · similarity 0.88)
- Safe and Stable Control via Lyapunov-Guided Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Multi-Robot Motion Planning with Diffusion Models (ICLR 2025 · similarity 0.87)
- Resisting Stochastic Risks in Diffusion Planners with the Trajectory Aggregation Tree (ICML 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.