End-to-end diffusion planning has shown strong potential for autonomous driving, but the physical feasibility of generated trajectories remains insufficiently addressed. In particular, generated trajectories may exhibit local geometric irregularities, violate trajectory-level kinematic constraints, or deviate from the drivable area, indicating that the commonly used noise-centric formulation in diffusion planning is not yet well aligned with the trajectory space where feasibility is more naturally characterized. To address this issue, we propose FeaXDrive, a feasibility-aware trajectory-centric diffusion planning method for end-to-end autonomous driving. The core idea is to treat the clean trajectory as the unified object for feasibility-aware modeling throughout the diffusion process. Built on this trajectory-centric formulation, FeaXDrive integrates adaptive curvature-constrained training to improve intrinsic geometric and kinematic feasibility, drivable-area guidance within reverse diffusion sampling to enhance consistency with the drivable area, and feasibility-aware GRPO post-training to further improve planning performance while balancing trajectory-space feasibility. Experiments on the NAVSIM benchmark show that FeaXDrive achieves strong closed-loop planning performance while substantially improving trajectory-space feasibility. These findings highlight the importance of explicitly modeling trajectory-space feasibility in end-to-end diffusion planning and provide a step toward more reliable and physically grounded autonomous driving planners.
FeaXDrive builds a unified feasibility-aware pipeline across training, inference, and post-training in clean trajectory space.
We propose a trajectory-centric diffusion planning framework with feasibility-aware training and inference for autonomous driving. By treating the clean trajectory as the central object for feasibility-aware modeling throughout the diffusion process, the framework provides a unified interface for training-time feasibility enhancement and inference-time geometric guidance in trajectory space.
We introduce an adaptive differentiable curvature training strategy to improve the intrinsic feasibility of generated trajectories. By directly imposing adaptive curvature constraints on the predicted clean trajectory, the method suppresses local geometric irregularities and curvature spikes while improving trajectory-level kinematic feasibility.
We develop a drivable-aware sampling guidance strategy for diffusion sampling to improve drivable-area consistency during inference. By injecting local road-geometry priors into the clean trajectory at each reverse sampling step, the method enables scene-aware geometric correction within the sampling loop.
We further incorporate feasibility-aware Group Relative Policy Optimization (GRPO) fine-tuning into the end-to-end diffusion planner. By optimizing a feasibility-augmented reward that jointly captures benchmark-oriented performance and trajectory-space feasibility preference, this stage further improves planning performance while balancing trajectory-space feasibility.
The predicted clean trajectory is the shared object for feasibility-aware modeling in training, inference, and post-training.
During training, FeaXDrive directly imposes adaptive differentiable curvature constraints on the predicted clean trajectory in trajectory space. This improves intrinsic geometric regularity, suppresses curvature spikes, and enhances trajectory-level kinematic feasibility.
During reverse diffusion sampling, the current clean trajectory estimate is guided with local road-geometry priors. Using a footprint-level drivable-area objective, the method performs online progressive correction so that the trajectory stays better aligned with the drivable region.
In post-training, FeaXDrive introduces feasibility-aware GRPO to optimize the planner with rewards that jointly reflect benchmark performance and trajectory-space feasibility, further improving the performance-feasibility trade-off.
Qualitative comparison between the noise-centric baseline and FeaXDrive on representative planning scenes.
From top to bottom, the representative cases show local geometric irregularities, trajectory-level kinematic infeasibility, and drivable-area inconsistency. Compared with the baseline, FeaXDrive produces trajectories that are smoother, more kinematically feasible, and better aligned with the drivable area.
If you find this project useful, please cite:
@article{wang2026feaxdrive,
title={FeaXDrive: Feasibility-aware Trajectory-Centric Diffusion Planning for End-to-End Autonomous Driving},
author={Wang, Baoyun and Li, Zhuoren and Liu, Ming and Zhang, Xinrui and Leng, Bo and Xiong, Lu},
journal={arXiv preprint arXiv:2604.12656},
year={2026}
}