ReflexDiffusion: Reflection-Enhanced Trajectory Planning for High-lateral-acceleration Scenarios in Autonomous Driving

TL;DR

ReflexDiffusion tackles the challenge of safe autonomous driving in long-tail, high-lateral-acceleration maneuvers by integrating an inference-time, physics-aware reflection into diffusion-based trajectory planning. The method injects gradient-based corrections that amplify the curvature-speed-acceleration coupling $a_y = \kappa v^2$ during diffusion sampling, removing the need for hand-crafted classifier guidance and remaining architecture-agnostic. Key contributions include a conditional dropout-based training strategy to simulate partial observability, a CFG-based denoising module, a reflection mechanism that projects gradient corrections onto the centripetal-force manifold, and a trajectory-confidence-driven trigger for safe refinement. Experiments on nuPlan Test14-hard and Test14-random demonstrate state-of-the-art improvements in high-risk scenarios (e.g., $14.1\%$ driving-score gain in reactive mode) while maintaining real-time performance ($\approx 36.1$ ms average) and preserving planning frequency, highlighting practical impact for robust autonomous driving in challenging conditions. Overall, ReflexDiffusion offers a general, inference-stage enhancement that can be deployed atop existing diffusion planners to substantially improve safety near handling limits without retraining.

Abstract

Generating safe and reliable trajectories for autonomous vehicles in long-tail scenarios remains a significant challenge, particularly for high-lateral-acceleration maneuvers such as sharp turns, which represent critical safety situations. Existing trajectory planners exhibit systematic failures in these scenarios due to data imbalance. This results in insufficient modelling of vehicle dynamics, road geometry, and environmental constraints in high-risk situations, leading to suboptimal or unsafe trajectory prediction when vehicles operate near their physical limits. In this paper, we introduce ReflexDiffusion, a novel inference-stage framework that enhances diffusion-based trajectory planners through reflective adjustment. Our method introduces a gradient-based adjustment mechanism during the iterative denoising process: after each standard trajectory update, we compute the gradient between the conditional and unconditional noise predictions to explicitly amplify critical conditioning signals, including road curvature and lateral vehicle dynamics. This amplification enforces strict adherence to physical constraints, particularly improving stability during high-lateral-acceleration maneuvers where precise vehicle-road interaction is paramount. Evaluated on the nuPlan Test14-hard benchmark, ReflexDiffusion achieves a 14.1% improvement in driving score for high-lateral-acceleration scenarios over the state-of-the-art (SOTA) methods. This demonstrates that inference-time trajectory optimization can effectively compensate for training data sparsity by dynamically reinforcing safety-critical constraints near handling limits. The framework's architecture-agnostic design enables direct deployment to existing diffusion-based planners, offering a practical solution for improving autonomous vehicle safety in challenging driving conditions.

Figures (4)

• Figure 1: Architecture of ReflexDiffusion. (a) Training Module enhances model robustness to physically-coupled features via a conditional dropout strategy. (b) Denoising Module employs classifier-free guidance to generate initial trajectories. (c) Reflection Module iteratively refines trajectories by injecting physics-aware gradients to enforce curvature-speed-acceleration coupling constraints during inference. (d) Trajectory Confidence Module assesses trajectory reliability, and dynamically triggers the reflection mechanism to ensure safety.
• Figure 2: Visualization comparison in U-turn scenario from high-lateral-acceleration scenarios. The Diffusion Planner's trajectory veers out of the lane during the turn, while ReflexDiffusion makes the turn without incident!
• Figure 3: Visualization of trajectory confidence in inference process. represents the trend of trajectory confidence for ReflexDiffusion and represents the trend of trajectory confidence for Diffusion Planner.
• Figure 4: Ablation Studies. The optimal values for the four parameters: conditional dropout rate, denoising scale $\lambda_1$, reflection scale $\lambda_2$ and confidence threshold $\gamma$ are 0.1, 0.9, 0.0, and 0.8, respectively.