Research map / Image generation / Fast diffusion sampling
DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving
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 DiffusionDrive, a truncated diffusion policy that denoises from anchored Gaussian noise around K-Means driving anchors with a cascade transformer diffusion decoder, achieving 88.1 PDMS on NAVSIM at 45 FPS.
Problem
Vanilla diffusion policies for end-to-end driving need many denoising steps (too slow for real time) and suffer mode collapse, producing overlapping trajectories in dynamic open-world traffic.
Approach
Truncate the diffusion schedule and start denoising from Gaussian noise centered on K-Means prior trajectory anchors (anchored Gaussian distribution), so only 2 denoising steps are needed, combined with a cascade diffusion decoder using deformable spatial cross-attention over BEV/PV features and agent/map queries.
Key results
- Achieves 88.1 PDMS on NAVSIM navtest with ResNet-34, versus 86.5 for Hydra-MDP-V8192-W-EP and 84.0 for Transfuser.
- Reduces denoising steps from 20 (TransfuserDP) to 2, running at 45 FPS on an NVIDIA 4090 versus 7 FPS for TransfuserDP.
- On nuScenes open-loop, achieves 0.57 average L2 and 0.08% average collision rate, with 1.8x speed of VAD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Transfuser regression with vanilla diffusion policy | Transfuser · NAVSIM navtest | PDMS: 84.0 -> 84.6 | Helped |
| use truncated diffusion policy | Transfuser · NAVSIM navtest | PDMS: 84.6 -> 85.7 | Helped |
| add the DiffusionDrive diffusion decoder to truncated diffusion | TransfuserTD · NAVSIM navtest | PDMS: 85.7 -> 88.1 | Helped |
| use DiffusionDrive instead of Hydra-MDP-V8192-W-EP | NAVSIM navtest · NAVSIM navtest | PDMS: 86.5 -> 88.1 | Helped |
| use DiffusionDrive instead of VAD | nuScenes · nuScenes | L2 (m), Avg.: 0.72 -> 0.57 | Helped |
| use DiffusionDrive instead of VAD | nuScenes · nuScenes | Collision Rate (%), Avg.: 0.22 -> 0.08 | Helped |
| remove spatial cross-attention (ablation) | DiffusionDrive · NAVSIM navtest | PDMS: 88.1 -> 55.1 | Hurt |
| add spatial cross-attention (ablation) | DiffusionDrive decoder with ego-query interaction · NAVSIM navtest | PDMS: 55.1 -> 87.1 | Helped |
| add agent/map cross-attention (ablation) | DiffusionDrive decoder with spatial cross-attention · NAVSIM navtest | PDMS: 87.1 -> 87.4 | Helped |
| add the cascade decoder (ablation) | DiffusionDrive decoder with spatial and agent/map cross-attention · NAVSIM navtest | PDMS: 87.4 -> 88.1 | Helped |
| use 1 denoising step instead of 2 (ablation) | DiffusionDrive · NAVSIM navtest | PDMS: 88.1 -> 87.9 | Hurt |
| use 3 denoising steps instead of 2 (ablation) | DiffusionDrive · NAVSIM navtest | PDMS: 88.1 -> 88.1 | No clear effect |
Surprising findings
- Denoising from anchored Gaussian with a single step already achieves 87.9 PDMS, indicating the anchors provide a strong starting point.
Limitations
- Evaluated on a limited set of driving benchmarks (NAVSIM, nuScenes); nuScenes scenarios are simple and only open-loop metrics used.
- Relies on K-Means anchors clustered from training data, so anchor quality depends on the dataset distribution.
Experimental setup
Datasets: NAVSIM navtrain nuScenes
Benchmarks: NAVSIM navtest nuScenes
Metrics: PDMS NC DAC TTC Comfort EP L2 error collision rate FPS mode diversity score
Baselines: Transfuser TransfuserDP UniAD VAD VADv2-V8192 Hydra-MDP-V8192-W-EP SparseDrive DRAMA PARA-Drive LTF
Builds on: diffusion policy DDIM Transfuser VADv2 SparseDrive
Closest papers by meaning
- Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 2024 · similarity 0.91)
- SceneDiffuser: Efficient and Controllable Driving Simulation Initialization and Rollout (NeurIPS 2024 · similarity 0.89)
- Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising (NeurIPS 2025 · similarity 0.88)
- Prediction with Action: Visual Policy Learning via Joint Denoising Process (NeurIPS 2024 · similarity 0.88)
- Prior Does Matter: Visual Navigation via Denoising Diffusion Bridge Models (CVPR 2025 · similarity 0.88)
- One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 2025 · similarity 0.88)
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.