Autonomous driving drops the correction loop: reflective guidance folded into one-pass inference, training time cut from 5.3 to 1.3 hours

A vision-language-action model for autonomous driving can now produce a safe trajectory in one forward pass with no external evaluation feedback, because the predict-evaluate-correct loop is compressed into the model's own parameters during training. Previously such models either optimized against a single scalar trajectory reward, which flags a poor trajectory but cannot say whether the error lies in scene understanding, behavior selection, or trajectory execution, nor how to fix it, or they had to run an extra prediction-evaluation-refinement loop at inference. On NAVSIM v1 the 0.8B, 2B, and 4B models reach PDMS of 89.3, 90.0, and 91.5, beating their corresponding reinforcement-learning baselines by 0.5, 0.8, and 0.3 points, with the 4B version taking the best overall PDMS in that table.

The mechanism has one model play student and teacher under different input contexts: the student sees only the standard input available at deployment, the teacher additionally sees reflective guidance built from evaluation results, and the student has its own generated trajectories scored token by token by the teacher, learning the teacher's more conservative choices. Reflective guidance is assembled from four templated components, a trajectory verdict, a planning diagnosis, an expert alignment, and a refinement directive, which turn scalar evaluation results and geometric differences from a reference trajectory into structured, trajectory-specific corrective supervision without another vision-language model or manual annotation. Training proceeds in three stages: driving-knowledge pretraining, planning-and-refinement supervised fine-tuning, then self-refinement rejection sampling, which keeps only refinements from the model's own rollouts that the evaluator verifies as better than the initial trajectory, narrowing the gap between offline refinement data and on-policy trajectories.

Distillation routes by trajectory quality: the evaluation score only decides which learning path a trajectory takes, with reliable trajectories kept by a self-preservation objective and risky ones receiving the teacher's token-level corrective distribution. Reinforcement learning collapsed trajectory-level reward into one value, so different failure causes received similarly low scores and repeated low-reward rollouts made policy optimization inefficient in long-tail scenarios. On NAVSIM v1 the guidance-conditioned teacher beats the unguided student by 6.1, 5.3, and 4.8 PDMS, and after distillation the student gains 3.0, 2.6, and 3.1 PDMS over its original performance, substantially closing the gap; per 10,000 samples, training time for the 0.8B model falls from 5.3 hours under GRPO to 1.3 hours under OPSD, with 2B and 4B needing only 1.7 and 2.4 hours.

The report covers a training framework and its benchmark evaluation: the model is tested for trajectory generation on NAVSIM v1/v2, two simulation-based planning benchmarks, closed-loop driving is evaluated on Bench2Drive, and open-loop L2 error and collision rate are reported on nuScenes, with training using Qwen3.5-0.8B/2B/4B on 8 NVIDIA H20 GPUs. It is aimed at researchers and engineers working on VLA post-training and distillation for autonomous driving. The source states no next-step experiment plan.

Self-refinement rejection sampling alone slightly lowers PDMS from 88.5 to 88.4, which the source attributes to partial forgetting of base planning ability during further training on refined samples; only combined with OPSD does it raise the score from 90.2 to 91.5. Raising the routing threshold from 0.65 to 0.85 gradually improves PDMS, but at 0.95 it falls back to 91.2, indicating that distilling many already reliable trajectories may disrupt effective student behaviors. Increasing the distillation support from 16 to 64 raises PDMS from 90.2 to 91.5, while 256 or the full vocabulary yields only marginal gains.