FeaXDrive: Feasibility-aware Trajectory-Centric Diffusion Planning for End-to-End Autonomous Driving

Abstract: 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.

• The paper introduces a trajectory-centric diffusion planning framework that embeds physical feasibility during the entire planning process for autonomous driving.
• It employs adaptive curvature-constrained training with dynamic loss formulation, significantly reducing curvature violation rates compared to baseline methods.
• FeaXDrive integrates local drivable-area guidance and feasibility-aware GRPO to boost benchmark metrics and achieve safe, physically-realizable trajectory generation.

FeaXDrive: Feasibility-Aware Trajectory-Centric Diffusion Planning for End-to-End Autonomous Driving

Introduction

The prevalence of end-to-end (E2E) planning frameworks in autonomous driving has been accelerated by the integration of vision-LLMs (VLMs) and generative models, notably diffusion models. While diffusion-based planners can capture the multi-modal characteristics and inherent uncertainty in driving scenarios, state-of-the-art methods inadequately constrain the generated trajectories to ensure physical executability. Notably, noise-centric parameterizations impede direct control over geometric and kinematic feasibility, leading to irregular, unfeasible, or undrivable paths. "FeaXDrive: Feasibility-aware Trajectory-Centric Diffusion Planning for End-to-End Autonomous Driving" (2604.12656) proposes a comprehensive trajectory-centric diffusion framework that explicitly embeds feasibility into the planning pipeline, offering both theoretical and practical advancements.

Trajectory-Centric Diffusion Planning

FeaXDrive adopts a trajectory-centric reformulation, where the clean, future trajectory serves as the central optimization target throughout the diffusion process. This contrasts with noise-centric models, whose prediction objectives lie primarily in the noise space—resulting in weaker and less interpretable feasibility signals. By operating directly in the trajectory space, FeaXDrive enables precise application of geometric and kinematic constraints during both training and inference, and streamlines supervision and correction pathways.

This trajectory-centric approach facilitates the direct encoding of local geometric regularity, speed-adaptive curvature constraints, and consistency with road geometry, directly tying the generative objectives to physically realizable driving trajectories. Experimental ablations confirm that simply shifting from noise-centric to trajectory-centric modeling yields reductions in both curvature and drivable-area violations, along with improvements in planning score metrics.

Adaptive Curvature-Constrained Training

To ensure that planned trajectories adhere to both geometric regularity and vehicle dynamics, the model introduces adaptive curvature-constrained training. Curvature estimation is implemented via differentiable smoothing and arc-length parameterization, which is robust to discretization artifacts and trajectory sampling rates.

In place of static geometric thresholds, a speed-dependent curvature bound is imposed. The model penalizes curvature excursions above the minimum of a fixed geometric threshold (matched to empirical vehicle minima, e.g., Chrysler Pacifica's turning radius) and a dynamic bound based on allowable lateral acceleration:

$$K_{\mathrm{adp}} = \min(K_{\max}, \frac{a_{\mathrm{lat}}}{v^2 + \epsilon_v})$$

This dynamic loss formulation ensures compliance across varying speed regimes, discouraging both low-speed artifacts (e.g., discrete spikes) and high-speed violations (excessive steer for velocity). Quantitative results show that this mechanism reduces curvature violation rates by more than an order of magnitude compared to baseline methods.

Constraint-Aware Inference via Drivable-Area Guidance

FeaXDrive introduces local drivable-area guidance within the reverse diffusion process. Utilizing geometric priors constructed as signed distance fields (SDFs) from high-definition (HD) maps, feasibility correction occurs at each sampling step, rather than as a post-hoc filter.

Critically, the guidance operates at the vehicle footprint level rather than the trajectory center, interpolating SDF values for all key contact points on the actual vehicle geometry. Soft barrier functions and gradient-based updates—triggered only when the predicted trajectory approaches boundaries or leaves the drivable area—modulate the clean trajectory estimates in situ without destabilizing otherwise feasible generation.

This guidance mechanism is lightweight in terms of compute and empirically leads to substantial improvements in drivable-area compliance—with DAC (drivable-area compliance) rates exceeding those of all competitive diffusion-based and non-diffusion baselines under identical benchmarks.

Feasibility-Aware GRPO Post-Training

The framework extends beyond imitation learning (IL) with Feasibility-Aware Group Relative Policy Optimization (GRPO), which integrates curvature and drivable-area feasibility directly into the post-training reward signal. Unlike conventional GRPO, which optimizes only for benchmark task metrics, this approach shapes the reward landscape to favor trajectories that jointly optimize for benchmark efficacy and feasibility metrics.

Reward shaping aggregates both PDMS and curvature- or boundary-feasibility bonuses; explicit within-batch comparison ensures that the optimization process does not regress on feasibility while maximizing score. Comparative experiments reveal that standard RLFT (reinforcement learning fine-tuning) often degrades feasibility for marginal score gains, whereas feasibility-aware GRPO preserves or improves both.

Empirical Evaluation

The method is assessed on the NAVSIM closed-loop autonomous driving benchmark, leveraging rich multi-modal VLM scene embeddings and high-level navigation cues.

Highlights:

• Under IL, FeaXDrive achieves a PDMS of 88.7, surpassing DiffusionDrive, WoTE, and other leading IL diffusion methods.
• Under joint IL and RLFT, PDMS is further improved to 90.0, with a DAC of 98.3.
• Curvature violation rates are reduced to 0.88% (IL) and 2.40% (IL+fa-GRPO), dramatically lower than competitive methods (DiffusionDrive: 8.59%).
• Drivable-area violation rates are similarly minimized, validating the effectiveness of joint feasibility-aware modeling across all stages.

Latency profiling demonstrates that the additional computational overhead for SDF construction and online feasibility guidance is insignificant relative to backbone VLM inference and conventional planning, supporting the method's deployability.

Theoretical Implications and Future Directions

The approach establishes that trajectory-space feasibility can—and should—be embedded directly into both generative modeling and downstream policy optimization. By aligning the generative data manifold and optimization objectives with physical constraints, the framework provides higher assurance of safety and regulatory compliance without sacrificing performance.

There remain opportunities for further integration, such as unifying more aspects of feasibility (beyond curvature and drivable area) within reward shaping, relaxing map dependence via online estimators, and advancing the efficiency of VLM backbones. Model generalization across map representations or under map uncertainty is a critical future avenue, as is the design of hierarchical constraint representations for multi-agent scenarios.

Conclusion

FeaXDrive demonstrates that explicit, trajectory-centric feasibility integration at all stages of the planner produces trajectories that simultaneously meet physical, geometric, and benchmark standards, addressing a pivotal shortcoming in generative E2E autonomous driving pipelines. The unified framework of adaptive constraints, streaming inference guidance, and policy-level reward shaping establishes a foundation for increasingly reliable and physically-consistent autonomous driving policies (2604.12656).