Source-linked AI summary
GoalFlow: Goal-Driven Flow Matching for Multimodal Trajectories Generation in End-to-End Autonomous Driving
Zebin Xing, Xingyu Zhang, Yang Hu, Bo Jiang, Tong He, Qian Zhang, Xiaoxiao Long, Wei Yin
TL;DR
GoalFlow addresses low-quality multimodal trajectories caused by divergent guidance and scene information in end-to-end autonomous driving. It uses scored goal points, flow matching, and trajectory selection, achieving state-of-the-art Navsim performance while retaining strong results with a single denoising step.
Problem
Existing end-to-end methods often emphasize collision and L2 metrics while overlooking drivable-area adherence, and deviating guidance can produce low-quality trajectories.
Method
GoalFlow scores candidate goal points using scene information, generates multimodal trajectories with flow matching and goal guidance, then scores trajectories for selection.
Results
GoalFlow achieves state-of-the-art results on Navsim, with a single denoising step reducing the score by only 1.6% from the optimal case.
Takeaways & Limitations
Goal-point guidance constrains multimodal trajectory generation, while flow matching supports efficient inference with fewer denoising steps.
Abstract
from arXiv · showhide
We propose GoalFlow, an end-to-end autonomous driving method for generating high-quality multimodal trajectories. In autonomous driving scenarios, there is rarely a single suitable trajectory. Recent methods have increasingly focused on modeling multimodal trajectory distributions. However, they suffer from trajectory selection complexity and reduced trajectory quality due to high trajectory divergence and inconsistencies between guidance and scene information. To address these issues, we introduce GoalFlow, a novel method that effectively constrains the generative process to produce high-quality, multimodal trajectories. To resolve the trajectory divergence problem inherent in diffusion-based methods, GoalFlow constrains the generated trajectories by introducing a goal point. GoalFlow establishes a novel scoring mechanism that selects the most appropriate goal point from the candidate points based on scene information. Furthermore, GoalFlow employs an efficient generative method, Flow Matching, to generate multimodal trajectories, and incorporates a refined scoring mechanism to select the optimal trajectory from the candidates. Our experimental results, validated on the Navsim\cite{Dauner2024_navsim}, demonstrate that GoalFlow achieves state-of-the-art performance, delivering robust multimodal trajectories for autonomous driving. GoalFlow achieved PDMS of 90.3, significantly surpassing other methods. Compared with other diffusion-policy-based methods, our approach requires only a single denoising step to obtain excellent performance. The code is available at https://github.com/YvanYin/GoalFlow.
1. Introduction
GoalFlow addresses low-quality and divergent multimodal trajectories by constraining generation with scene-informed goal points and flow matching, then selecting reliable candidates. Experiments on Navsim report state-of-the-art performance and strong single-step inference efficiency.
- Existing end-to-end systems often overlook whether generated vehicles remain within the drivable area and can produce low-quality trajectories when guidance deviates from ground truth.
- GoalFlow selects a goal point that is closest to the ground-truth goal and lies within a drivable area, providing stronger trajectory guidance.
- GoalFlow uses flow matching with scene information and selected-goal guidance to generate multimodal trajectories, followed by trajectory scoring.
- Navsim experiments show that GoalFlow outperformed other approaches in overall scoring and achieved a 1.6% score drop with only one denoising step versus the optimal case.
- GoalFlow introduces a trajectory-selection mechanism using shadow trajectories to address potential goal-point errors.
2. Related Work
Related work spans multimodal end-to-end driving, diffusion-based generation, and flow matching. GoalFlow builds on these directions by combining multimodal trajectory generation with a structured goal-driven architecture.
- End-to-End Autonomous Driving: End-to-end driving methods progressively fused camera and LiDAR features and propagated shared perception features into prediction and planning.
- Diffusion Models and Flow Matching: Diffusion models generate outputs by iteratively adding and removing noise, while flow matching learns a vector field that transitions between distributions.
- Goal-Driven Generation: GoalFlow organizes perception, goal-point construction, trajectory planning, and trajectory scoring into an architecture for guided multimodal generation.
- Multimodal Trajectory Generation: Planning scenarios often contain multiple valid actions, motivating multimodal modeling through discrete commands or other guidance signals.
3. Method
GoalFlow combines multimodal perception, goal-point guidance, and flow matching to generate and select constrained multimodal trajectories. Its goal-point scorer favors candidates close to the ground-truth endpoint and inside the drivable area, while Rectified Flow provides an efficient generative process.
- 3.1. Preliminary: Flow Matching learns a direction field that maps a simple noise distribution to the target trajectory distribution along straight-line paths.Rectified Flow uses xt = (1 − t)x0 + tx1 and trains a neural network to predict x1 − x0.
- 3.2.1. Overview: GoalFlow fuses camera and LiDAR features into a BEV representation for scene-aware trajectory generation.The perception module follows Transfuser, using separate backbones and multilayer transformer fusion to form Fbev.
- 3.2.3. Goal Point Construction Module: The method constructs a dense goal-point vocabulary by clustering trajectory endpoints, typically using 4096 or 8192 cluster centers.Each goal point represents position and heading, providing candidate guidance without relying on lane-level HD-map information.
- 3.2.3. Goal Point Construction Module: Each candidate goal point is scored for distance to the ground-truth endpoint and drivable-area compliance, then the highest final score determines the guidance point.A shadow vehicle tests whether the candidate’s footprint lies within the drivable polygon; the scorer decoder predicts the distance and compliance scores from scene features.
- 3.2.4. Trajectory Planning Module: The trajectory planner conditions flow matching on scene, ego, trajectory, time, and goal-point features to generate multimodal trajectory candidates.The model reconstructs trajectories by transforming noise toward normalized target trajectories and then denormalizing the result.
4. Experiments
GoalFlow is evaluated on Navsim using multimodal trajectory generation, goal-point guidance, trajectory scoring, and flow matching. It consistently outperforms competing methods, while ablations show benefits from goal-point guidance, multiple evaluators, trajectory scoring, model scaling, and efficient inference.
- Experimental Setup: GoalFlow is evaluated on the Openscene dataset through Navsim, using 1192 trainval scenarios and 136 test scenarios with multimodal sensor and annotation inputs.Each sample contains eight camera perspectives, fused LiDAR from five sensors, ego status, and map and object annotations.
- Sensitivity Analysis: Noise variance materially affects performance: scores drop significantly when σ exceeds 0.1 but remain stable below 0.1.At σ = 0.3, the Comfort score drops to 2.23 because generated trajectories become excessively erratic.
- Comparison with SOTA Methods: GoalFlow consistently outperforms other methods overall, surpassing the second-best approach by 5.5 DAC points and 5.7 EP points.Using the ground-truth trajectory endpoint as the goal point yields a score of 92.1, close to the human trajectory score of 94.8.
- Ablation Studies: Goal-point guidance produces the largest ablation improvement by decomposing planning into goal-point prediction and goal-guided trajectory generation.M1 extends the rectified-flow base model by selecting the highest-scoring point from a distance-score distribution.
- Ablation Studies: Adding DAC-score prediction and trajectory scoring further improves assessment robustness and GoalFlow reliability.M2 introduces multiple evaluators, while M3 adds trajectory selection and goal-point checking.
- Inference Efficiency: Inference remains effective with one denoising step, reducing one-sample denoising time to 6% of the original.Performance remains stable as inference steps decrease from 20 to 1, reflecting rectified flow’s direct path between distributions.
- Scaling Analysis: Increasing the Transformer hidden dimension and using a stronger image backbone improve end-to-end performance, with the best hidden dimension reported as 1024.Scaling the goal-point construction module also significantly improves overall performance.
5. Conclusion
GoalFlow targets accurate and efficient multimodal trajectory generation by combining precise goal points with constrained generative modeling. Experiments in Navsim report state-of-the-art performance, while future work will examine alternative guidance information.
- Conclusion: GoalFlow generates accurate and efficient multimodal trajectories by constraining a generative model with precise goal points.The framework is intended to produce high-quality multimodal trajectories.
- Conclusion: Experiments in the Navsim environment demonstrate state-of-the-art performance for GoalFlow.
- Future Work: Future work will investigate how different guidance information affects multimodal trajectory generation.