Research map / Image generation / Autoregressive and latent generation

SAH-Drive: A Scenario-Aware Hybrid Planner for Closed-Loop Vehicle Trajectory Generation

ICML 2025 · Image generation › Autoregressive and latent generation · approach: Diffusion planning and policies

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 SAH-Drive, a hybrid planner that switches between PDM-Closed (rule-based) and a lightweight diffusion proposal generator (learning-based) using a dual-timescale decision neuron with an STDP-inspired rule, reaching SOTA on interPlan.

Problem

Autonomous driving planners face a trade-off: rule-based planners are efficient but generalize poorly to long-tail scenarios, while learning-based planners generalize better but are slow and less interpretable.

Approach

A scenario-aware hybrid paradigm that runs a rule-based planner (PDM-Closed) for regular driving and a diffusion proposal generator for long-tail cases, with a dual-timescale decision neuron (score-based rule, scenario-based rule, and score-based STDP decision neuron) to switch planners, plus a dynamic diffusion proposal number regulator and exponentially weighted trajectory fusion.

Key results

• SAH-Drive achieves an interPlan score of 64, above Diffusion-ES (57) and PDM-Closed (42), the highest among compared planners.
• On Test14-Hard (R), SAH-Drive scores 83 versus 77 for Diffusion-ES and 75 for PDM-Closed.
• The learning-based planner is active for 6.70% of time in regular scenarios and 43.33% in long-tail scenarios; SAH-Drive's runtime is up to about 1 s per frame versus 0.1 s for PDM-Closed.

Measured effects reported in the paper

Surprising findings

• A diffusion model trained only on nuPlan Mini outperforms models trained on the full nuPlan dataset, and interPlan score drops as training data grows (65 on Mini vs 59 with Singapore and Boston).
• Using only the STDP-based decision neuron gives a smaller performance drop than using only the score-based or scenario-based switching rules.

Limitations

• Diffusion-sampled trajectories do not fully conform to real-world physical dynamics and serve only as proposals.
• Computational cost of learning-based component still up to ~1 s per frame.
• Performance on interPlan decreases with larger training datasets, suggesting capacity or distribution mismatch.

Experimental setup

Datasets: nuPlan Mini nuPlan Singapore nuPlan Boston

Benchmarks: interPlan Val14 Test14-Random Test14-Hard

Metrics: CLS-R CLS-NR PDM score

Baselines: PDM-Closed Diffusion-ES STR2 HybridLLMPlanner PlanTF Pluto Diffusion Planner IDM

Builds on: PDM-Closed Diffusion-ES STDP

Closest papers by meaning

• Diffusion-Based Planning for Autonomous Driving with Flexible Guidance (ICLR 2025 · similarity 0.88)
• Hybrid-Driving: An Autonomous Driving Decision Framework Integrating Large Language Models, Knowledge Graphs and Driving Rules (AAAI 2025 · similarity 0.86)
• On the Road to Portability: Compressing End-to-End Motion Planner for Autonomous Driving (CVPR 2024 · similarity 0.85)
• AlignDrive: Aligned Lateral-Longitudinal Planning for End-to-End Autonomous Driving (NeurIPS 2026 · similarity 0.85)
• Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.84)
• Future-Aware End-to-End Driving: Bidirectional Modeling of Trajectory Planning and Scene Evolution (NeurIPS 2025 · similarity 0.84)