Research map / Probabilistic and scientific modeling / Molecules and proteins
Flow Matching-Based Autonomous Driving Planning with Advanced Interactive Behavior Modeling
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 Flow Planner, a flow-matching transformer planner for autonomous driving that uses fine-grained overlapping trajectory tokenization, scale-adaptive spatiotemporal fusion, and classifier-free guidance that reweights neighboring-vehicle conditioning to better model interactive behavior.
Problem
Learning-based driving planners struggle to model interactive behaviors with neighboring agents, due to vanilla transformer fusion of heterogeneous scene information and scarce interactive training data for imitation learning.
Approach
Combine overlapping trajectory-segment tokens, distance-scaled (scale-adaptive) attention with separate adaLN/FFN per modality, and flow matching with classifier-free guidance that masks neighboring-vehicle conditions to amplify interaction cues at inference.
Key results
- Achieves 90.43 on nuPlan Val14 (non-reactive) without rule-based refinement, the first learning-based method reported above 90 on that benchmark per the authors.
- On interPlan, reaches 61.82 overall vs. 52.90 for Diffusion Planner (about 8.9-point gain).
- Ablation path raises Val14 from 88.10 (base) to 90.43 with the full model, with classifier-free guidance adding about 0.9 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add fine-grained trajectory tokenization (ablation) | Flow Planner base model · nuPlan Val14 and interPlan | score: 88.10 -> 88.33; 41.27 -> 44.14 | Helped |
| add fine-grained trajectory tokenization (ablation) | Flow Planner base model · interPlan Nudge Around | score: 50.61 -> 43.27 | Hurt |
| add scale-adaptive attention (ablation) | Flow Planner with trajectory tokenization · nuPlan Val14 and interPlan | score: 88.33 -> 88.77; 44.14 -> 46.25 | Helped |
| add separate adaLN and FFN modules (ablation) | Flow Planner with scale-adaptive attention · nuPlan Val14 and interPlan | score: 88.77 -> 89.54; 46.25 -> 58.22 | Helped |
| add classifier-free guidance (ablation) | Flow Planner with separate adaLN and FFN · nuPlan Val14 and interPlan | score: 89.54 -> 90.43; 58.22 -> 61.82 | Helped |
| increase classifier-free guidance scale, then increase beyond 1.80 (ablation) | Flow Planner · nuPlan Val14 | score: 89.64; 89.89; 90.14; 90.43; 90.00; 89.63 | Mixed |
| use Flow Planner instead of Diffusion Planner | learning-based planners · nuPlan Val14 NR | score: 89.87 -> 90.43 | Helped |
| use Flow Planner instead of Diffusion Planner | learning-based planners · interPlan Overall | score: 52.90 -> 61.82 | Helped |
Surprising findings
- Scale-adaptive attention alone gave no visible improvement; it only helped when combined with separate adaLN and FFN per modality.
Limitations
- Inference speed only marginally above 10 Hz due to spatiotemporal self-attention cost.
- Interactive modeling is implicit and still imitation-learning based, facing data quality and forgetting issues.
- Code is promised only upon acceptance; no release at submission.
Experimental setup
Datasets: nuPlan interPlan
Benchmarks: nuPlan Val14 nuPlan Test14-random nuPlan Test14-hard interPlan
Metrics: closed-loop planning score (NR) closed-loop planning score (R) overall score on interPlan
Baselines: IDM PDM-Closed PDM-Hybrid PLUTO GameFormer Diffusion Planner PlanTF PDM-Open
Builds on: flow matching classifier-free guidance scale-adaptive attention Diffusion Planner MLP-Mixer
Closest papers by meaning
- Diffusion-Based Planning for Autonomous Driving with Flexible Guidance (ICLR 2025 · similarity 0.88)
- Beyond Self-Play and Scale: A Behavior Benchmark for Generalization in Autonomous Driving (NeurIPS 2026 · similarity 0.88)
- Trajeglish: Traffic Modeling as Next-Token Prediction (ICLR 2024 · similarity 0.88)
- ChainFlow-VLA: Causal Flow Planning with Vision-Language Models (NeurIPS 2026 · similarity 0.88)
- UniMotion: A Unified Motion Framework for Simulation, Prediction and Planning (NeurIPS 2025 · similarity 0.87)
- Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 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.