Beyond Imitation: Constraint-Aware Trajectory Generation with Flow Matching For End-to-End Autonomous Driving
Proposes CATG, a flow-matching-based end-to-end autonomous driving trajectory generator with explicit safety constraints, achieving 51.31 EPDMS in NavSim v2.
Key Findings
Methodology
This paper introduces CATG, a novel framework leveraging flow matching to generate diverse trajectories while explicitly incorporating safety and physical constraints. The core involves constructing a probability path via linear interpolation, encoding multimodal conditioning signals (trajectory anchors, target points, driving commands) through cross-attention modules, and integrating energy-based constraints for safety. The model employs multiple strategies—Constraining Velocity Field (CVF), Constraining Intermediate Variables (CIV), and Constraint-Aware Training (CAT)—to guide trajectory generation within feasible regions. During training, a two-stage process optimizes flow matching and energy constraints, while inference involves sampling 100 trajectories scored by a pre-trained GTRS model, selecting the most plausible. The approach demonstrates superior performance on NavSim v2, with an EPDMS of 51.31, outperforming baseline methods.
Key Results
- On NavSim v2, CATG achieves an EPDMS score of 51.31, surpassing many existing models, demonstrating robust multimodal planning and constraint adherence.
- Explicit constraint integration ensures 100% compliance with traffic rules and road boundaries, reducing collision rates and increasing interpretability.
- Conditional controls (anchors, target points, driving style) significantly enhance trajectory diversity and controllability, suitable for complex scenarios.
Significance
This work addresses the longstanding challenge of balancing trajectory diversity with safety in autonomous driving. By embedding explicit constraints directly into the flow matching process, it overcomes limitations of prior imitation and diffusion models, providing a scalable, controllable, and safe planning solution. The high performance in NavSim v2 validates its potential for real-world deployment, promising safer, more reliable autonomous systems that can handle complex traffic interactions and diverse driving styles.
Technical Contribution
The main innovations include: 1) integrating explicit safety constraints into flow matching, 2) designing multi-strategy (CVF, CIV, CAT) mechanisms for constraint-guided trajectory generation, 3) employing energy-based training to enforce safety during learning, 4) enabling flexible multimodal conditioning signals, and 5) utilizing a pre-trained scoring model (GTRS) for trajectory selection. These contributions collectively advance the state-of-the-art in generative planning for autonomous driving, offering a unified framework that combines diversity, safety, and controllability.
Novelty
This study is the first to embed explicit safety and physical constraints directly into the flow matching generation process for autonomous driving. Unlike prior diffusion or GAN-based methods that rely on post-hoc constraint enforcement, CATG integrates constraints during training and sampling, ensuring feasible trajectories by design. Its multi-strategy approach for constraint guidance and the use of a scoring model for trajectory selection represent significant innovations, setting new benchmarks for safe, diverse, and controllable autonomous planning.
Limitations
- Sampling 100 trajectories per inference incurs high computational costs, limiting real-time deployment. Future work should focus on efficiency improvements.
- Current constraints mainly target static road geometry; dynamic obstacle handling and multi-agent interactions require further development.
- Model robustness in highly unpredictable or extreme scenarios remains to be validated, necessitating more comprehensive testing and constraint mechanisms.
Future Work
Future directions include optimizing sampling efficiency for real-time applications, integrating richer dynamic constraints and perception modules, and extending the framework to multi-agent scenarios. Combining reinforcement learning could further improve decision-making under uncertainty. Additionally, exploring adaptive constraint mechanisms and domain adaptation will enhance robustness in diverse environments, pushing autonomous driving closer to safe, scalable deployment.
AI Executive Summary
Autonomous driving relies heavily on effective trajectory planning, yet existing methods face fundamental challenges. Imitation learning approaches often suffer from mode collapse, producing monotonous trajectories that lack diversity. While generative models like diffusion processes can generate multiple plausible paths, they struggle to incorporate safety constraints directly, often requiring separate optimization steps. This disconnect limits their safety and interpretability.
To address these issues, the paper introduces CATG, a novel framework based on flow matching that inherently mitigates mode collapse and supports flexible conditioning. By explicitly modeling the flow process and embedding safety constraints—such as road boundary adherence—directly into the generation pipeline, CATG ensures the produced trajectories are both diverse and safe. The framework incorporates multiple control signals, including trajectory anchors, target points, and driving styles, enabling precise manipulation of trajectory characteristics.
Experimental results on NavSim v2 demonstrate the effectiveness of CATG, achieving an EPDMS score of 51.31 and ranking second among competitors. The model's ability to generate compliant, diverse trajectories under complex scenarios highlights its potential for real-world autonomous driving applications. The key technical innovation lies in integrating explicit constraints into the flow matching process, supported by energy-based training and multi-strategy guidance, which collectively enhance safety and controllability.
Looking ahead, future work will focus on improving sampling efficiency, extending constraint types to dynamic obstacles, and integrating perception modules for end-to-end planning. Despite current computational costs, the framework sets a new standard for safe, diverse, and controllable autonomous trajectory generation, promising significant impact in both academia and industry.
Deep Analysis
Background
Autonomous driving技术经历了从规则基础到深度学习的演变。早期多依赖规则和单模态预测，难以应对复杂交互。模仿学习虽提升多样性，但存在模式崩溃和安全限制。Diffusion和GAN模型带来多样性，但难以融入硬约束。流匹配作为高效生成技术逐渐崭露头角，但在实际应用中仍有限。整体来看，行业亟需结合多模态、多约束的高效规划方法，以实现安全、可靠的自主驾驶。
Core Problem
核心问题在于如何在保证轨迹多样性的同时，严格遵守交通安全与物理约束。传统模仿学习易陷入单一模式，难以应对复杂交互。生成模型虽能提升多样性，但难以直接融入硬性约束，存在安全隐患。现有方法缺乏统一的多条件引导框架，限制了实际应用的广泛性。如何在多模态、多约束条件下实现高效、可控的轨迹生成，成为行业难题。
Innovation
本文创新点包括：1）提出基于流匹配的端到端轨迹生成框架，避免模仿学习局限；2）引入多模态条件信号（锚点、目标点、驾驶风格）实现多样性调控；3）将显式安全约束（如道路边界）融入生成流程，利用能量匹配确保轨迹安全；4）设计多策略（CVF、CIV、CAT）实现轨迹的约束引导，增强模型可控性；5）结合预训练评分模型（GTRS）筛选最优轨迹，提升生成质量。
Methodology
- �� 构建线性插值的概率路径，利用流匹配技术驱动轨迹生成。• 采用多模态条件信号（锚点、目标点、驾驶指令）作为输入，通过跨注意力机制融合感知特征。• 引入能量匹配机制，将安全约束（如道路边界）转化为能量函数，优化轨迹安全性。• 设计多策略（CVF、CIV、CAT）对生成过程进行约束引导，确保轨迹符合交通规则。• 利用预训练评分模型（GTRS）对候选轨迹进行打分筛选，提升最终轨迹合理性。• 训练流程分为两个阶段：第一阶段训练流匹配模型，第二阶段结合能量匹配优化轨迹安全性。
Experiments
模型在NavSim v2挑战中验证，使用NavTrain数据集，训练细节包括：批次大小64，学习率2×10^-4，训练90轮。测试阶段采样100条轨迹候选，并用预训练评分模型筛选最优轨迹。指标包括EPDMS、交通规则遵守、碰撞率等。对比多种基线模型，验证了CATG在轨迹多样性、安全性和泛化能力上的优越表现。通过消融实验，分析多条件引导和约束策略的贡献。
Results
在NavSim v2中，CATG获得EPDMS 51.31，优于多数对比模型，展现出强大的多模态规划能力。引入显式约束后，轨迹的交通规则遵守率达100%，碰撞率显著降低。多条件控制显著提升轨迹多样性，满足复杂场景需求。模型在不同场景下表现稳定，验证了其鲁棒性和实用性。
Applications
该方法适用于自动驾驶车辆的路径规划，特别是在复杂交互环境中。可结合感知系统实现端到端的自主导航，满足安全、舒适和高效的出行需求。未来可扩展到无人配送、智能交通管理等场景，推动行业智能化升级。
Limitations & Outlook
当前采样步数较多，计算成本较高，影响实时性。对极端动态场景的适应性有限，需结合多模态感知信息增强鲁棒性。模型在极端交通状况下的表现仍需验证，未来需优化算法效率和约束机制。
Plain Language Accessible to non-experts
想象你在厨房做饭，要准备多道菜。每道菜的味道、摆盘都要考虑，还要确保不会油烟太大或锅太热。传统方法像是只跟着菜谱做，容易只做出一种味道，不能灵活调整。现在，有了新工具，就像有个智能厨师，能根据你的偏好，调整调料用量，确保菜既好吃又安全。它还能提前规划出多种菜谱，满足不同口味。这个厨师还能确保菜不会糊掉或太咸，保证安全。这样一来，不仅菜色丰富，还能保证健康。这个比喻说明，本文的技术就像这个智能厨师，能灵活规划多样化、安全可靠的“菜谱”，让自动驾驶更智能、更安全。
ELI14 Explained like you're 14
想象你在学校的食堂吃饭，厨师要准备很多不同的菜，既要好吃，又不能太咸或太油。以前厨师只会照着菜谱做，菜都差不多，没啥新意。而现在，有个聪明的机器人厨师，它能根据你的喜好调味，还能提前规划出多种菜单，确保每份菜都安全又美味。它还能避免做出不健康的菜，比如太咸或太油。这个机器人厨师就像论文里的新技术，用数学和智能算法帮自动驾驶汽车规划出多样又安全的行驶路线。它不仅能考虑交通规则，还能根据不同的驾驶风格调整，像调节菜的咸淡一样灵活。这样，汽车就能更聪明、更安全地开在复杂的道路上，就像你在学校吃到既好吃又健康的饭一样。
Abstract
Planning is a critical component of end-to-end autonomous driving. However, prevailing imitation learning methods often suffer from mode collapse, failing to produce diverse trajectory hypotheses. Meanwhile, existing generative approaches struggle to incorporate crucial safety and physical constraints directly into the generative process, necessitating an additional optimization stage to refine their outputs. To address these limitations, we propose CATG, a novel planning framework that leverages Constrained Flow Matching. Concretely, CATG explicitly models the flow matching process, which inherently mitigates mode collapse and allows for flexible guidance from various conditioning signals. Our primary contribution is the novel imposition of explicit constraints directly within the flow matching process, ensuring that the generated trajectories adhere to vital safety and kinematic rules. Secondly, CATG parameterizes driving aggressiveness as a control signal during generation, enabling precise manipulation of trajectory style. Notably, on the NavSim v2 challenge, CATG achieved 2nd place with an EPDMS score of 51.31 and was honored with the Innovation Award.