Research map / Image generation / Autoregressive and latent generation

Diffusion-Based Planning for Autonomous Driving with Flexible Guidance

ICLR 2025 · Image generation › Autoregressive and latent generation · approach: Diffusion planning and policies · 2 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 Diffusion Planner, a DiT-based transformer that jointly generates ego and neighbor trajectories via a diffusion objective and uses training-free classifier guidance for safe, adaptable closed-loop planning on nuPlan and a 200-hour delivery-vehicle dataset.

Problem

Learning-based autonomous driving planners struggle to model multi-modal human driving behavior, rely on rule-based refinement for safety, and are hard to adapt post-training to safety or personalized driving preferences.

Approach

Model planning and neighbor prediction jointly as a diffusion trajectory generation task with a DiT-style transformer, and inject safety/comfort/speed preferences at inference via differentiable energy-based classifier guidance, without retraining or rule-based post-processing.

Key results

• Achieves 89.19 non-reactive score on nuPlan Test14 and 82.93 reactive on Test14 among learning-based planners, outperforming PLUTO w/o refine. (89.90 NR) is close but PlanTF scores lower.
• With refinement module, Diffusion Planner w/ refine. reaches 94.80 NR on Test14, surpassing rule-based and hybrid baselines.
• On the delivery-vehicle dataset, Diffusion Planner scores 92.08 versus 90.89 for PlanTF and 87.77 for PLUTO w/o refine.
• Inference runs at about 20 Hz on a single A6000 GPU with 10 denoising steps.

Measured effects reported in the paper

Surprising findings

• Learning-based Diffusion Planner without refinement does not clearly beat PLUTO w/o refine. on the nuPlan Test14 non-reactive score (89.19 vs 89.90).
• Existing diffusion planners such as STR2-CPKS-800M and Diffusion-es degrade substantially without LLM filtering or rule-based refinement.

Limitations

• Relies on vectorized map and detected neighbor inputs rather than end-to-end sensor inputs.
• Weak lateral flexibility and lane-change behavior due to straight-driving-dominated data.
• Multiple denoising inference steps reduce sample efficiency compared with single-pass planners.

Experimental setup

Datasets: nuPlan delivery-vehicle driving dataset (200 hours)

Benchmarks: nuPlan Val14 nuPlan Test14 nuPlan Test14-hard delivery-vehicle driving dataset

Metrics: nuPlan closed-loop score collisions TTC drivable area compliance comfort progress

Baselines: IDM PDM-Closed PDM-Hybrid PDM-Open UrbanDriver GameFormer PlanTF PLUTO Diffusion-es STR2-CPKS-800M

Builds on: DiT classifier guidance DPM-Solver PDM diffusion posterior sampling

Closest papers by meaning

• Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 2024 · similarity 0.91)
• Flow Matching-Based Autonomous Driving Planning with Advanced Interactive Behavior Modeling (NeurIPS 2025 · similarity 0.88)
• SAH-Drive: A Scenario-Aware Hybrid Planner for Closed-Loop Vehicle Trajectory Generation (ICML 2025 · similarity 0.88)
• ChainFlow-VLA: Causal Flow Planning with Vision-Language Models (NeurIPS 2026 · similarity 0.88)
• DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving (CVPR 2025 · similarity 0.88)
• Beyond Self-Play and Scale: A Behavior Benchmark for Generalization in Autonomous Driving (NeurIPS 2026 · similarity 0.87)