Research map / Image generation / Fast diffusion sampling
GoalFlow: Goal-Driven Flow Matching for Multimodal Trajectories Generation in 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 GoalFlow, an end-to-end driving planner that selects a goal point from a dense vocabulary and uses rectified flow conditioned on it to generate multimodal trajectories, reaching PDMS 90.3 on NavSim.
Problem
Multimodal trajectory generation for end-to-end autonomous driving suffers from high trajectory divergence in diffusion models and low-quality trajectories when pre-defined guidance (goal points or commands) deviates from ground truth.
Approach
Decompose planning into goal point construction (scoring a dense goal-point vocabulary for distance-to-GT and drivable-area compliance using shadow vehicles, without HD maps) and goal-conditioned rectified flow generation, plus a trajectory scorer that selects among candidates using goal-point distance and progress, with shadow trajectories as a fallback.
Key results
- GoalFlow reaches PDMS 90.3 on NavSim test, above Transfuser (84.0) and PARA-Drive (84.0), with DAC 98.3.
- Using the ground-truth endpoint as goal point yields PDMS 92.1, close to the human score of 94.8.
- A single denoising step gives PDMS 88.9 versus 90.3 with 5 steps, while cutting denoising time to about 6% of the 20-step cost.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GoalFlow instead of PARA-Drive | Navsim autonomous driving · Navsim Test | SPDM: 84.0 -> 90.3 | Helped |
| add a distance score map to guide rectified flow (ablation) | GoalFlow M0 · Navsim Test | SPDM: 85.6 -> 88.5 | Helped |
| add a DAC score map (ablation) | GoalFlow M1 · Navsim Test | SPDM: 88.5 -> 89.4 | Helped |
| add the trajectory scorer (ablation) | GoalFlow M2 · Navsim Test | SPDM: 89.4 -> 90.3 | Helped |
| reduce inference steps from 20 to 1 (ablation) | GoalFlow · Navsim Test | SPDM: 89.9 -> 88.9 | Hurt |
| set initial noise standard deviation σ to 0.2 instead of 0.1 (ablation) | GoalFlow · Navsim Test | SPDM: 90.3 -> 49.0 | Hurt |
| set initial noise standard deviation σ to 0.3 instead of 0.1 (ablation) | GoalFlow · Navsim Test | SPDM: 90.3 -> 18.8 | Hurt |
| increase Transformer hidden dimension from 256 to 512 (ablation) | GoalFlow M2 · Navsim Test | SPDM: 86.5 -> 88.1 | Helped |
| increase Transformer hidden dimension from 512 to 1024 (ablation) | GoalFlow M2 · Navsim Test | SPDM: 88.1 -> 89.4 | Helped |
| replace V2-99 image backbone with ResNet-34 (ablation) | GoalFlow M2 · Navsim Test | SPDM: 86.5 -> 85.7 | Hurt |
Surprising findings
- Flow matching with a single inference step remains close to the best multi-step result, only about 1.4 PDMS points lower.
Limitations
- Goal point errors can misguide trajectories, requiring shadow-trajectory fallback.
- Relies on the goal-point vocabulary built from training endpoints; ablation shows sensitivity to initial noise sigma above 0.1.
Experimental setup
Datasets: OpenScene NavSim
Benchmarks: NavSim
Metrics: PDM score no at-fault collisions drivable area compliance time to collision ego progress comfort L2
Baselines: Constant Velocity Ego Status MLP PDM-Closed Transfuser LTF UniAD PARA-Drive
Builds on: rectified flow Transfuser VADv2 classifier-free guidance Diffusion-ES
Closest papers by meaning
- ChainFlow-VLA: Causal Flow Planning with Vision-Language Models (NeurIPS 2026 · similarity 0.88)
- Future-Aware End-to-End Driving: Bidirectional Modeling of Trajectory Planning and Scene Evolution (NeurIPS 2025 · similarity 0.87)
- Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 2024 · similarity 0.87)
- Generating Multimodal Driving Scenes via Next-Scene Prediction (CVPR 2025 · similarity 0.86)
- Flow Matching-Based Autonomous Driving Planning with Advanced Interactive Behavior Modeling (NeurIPS 2025 · similarity 0.86)
- Don't Shake the Wheel: Momentum-Aware Planning in End-to-End Autonomous Driving (CVPR 2025 · similarity 0.86)
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.