Beyond Imitation: Constraint-Aware Trajectory Generation with Flow Matching For End-to-End Autonomous Driving

TL;DR

The paper tackles the challenge of end-to-end autonomous driving planning where imitation learning yields limited diversity and purely generative methods lack hard constraints. It introduces CATG, a Constraint-Aware Trajectory Generation framework built on Flow Matching that eliminates imitation learning and supports flexible conditioning signals and explicit constraint integration. It leverages three constraint mechanisms—CVF, CIV, and CAT—together with Energy Matching in a two-stage training regime, plus reward-conditioned controllability via an EP score, all within a perception-driven conditioning pipeline using trajectory anchors, target points, and driving commands. On NAVSIM v2, CATG achieves a second-place EPDMS of 51.31 and an Innovation Award, demonstrating robust constraint-aware trajectory generation and controllable driving style for end-to-end autonomous planning.

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.

Figures (2)

• Figure 1: The architecture of CATG with Flow Matching. The image is encoded into image features, which are subsequently fed into detection and mapping modules to generate perceptual features. Before training, CATG processes each gt trajectory (GT Traj) through simulator to compute and offline store its score. Subsequently, CATG encodes Gaussian-sampled latent variables and the timestamp $t$ into input features via a encoder. These features are fused with perceptual features, trajectory rewards, and conditioning signals (including plan anchor, goal point, and driving command) through cross-attention. The signal fusion module consists of multi-layer cross-attention blocks and adheres to the classifier-free classifierfreediffusionguidance training paradigm. The velocity field decoded by the decoder is refined via our three correction strategies: CVF, CIV, and CAT (Sec \ref{['sec:CVF']}). Finally, the driving trajectory is generated through sampling.
• Figure 2: The figure illustrates that sampling by using an anchor as the starting point results in a more reasonable trajectory.