Organizations: Connected and Autonomous Vehicles Lab (CAV-Lab), University of Surrey, Guildford, GU2 7XH, UK. · Department of Mechanical and Aerospace Engineering, George Washington University, Washington, DC 20052, USA.
Vision-language-action (VLA) models offer a unified approach to perception, reasoning, and control for autonomous driving, but accurate action prediction alone does not establish compliance with traffic regulations. This paper presents a rule-guided VLA framework that incorporates symbolic traffic-rule feedback into policy optimisation. The framework combines supervised fine-tuning on expert demonstrations with reinforcement learning (RL) refinement using a reward that encourages encoded-rule compliance while penalising deviations from expert actions. Selected provisions of the United Kingdom Highway Code are formalised as executable first-order logic rules using PyReason, covering speed adaptation, following distance, and road-user priority. A frozen vision-language model extracts scene predicates, which are combined with vehicle-state information to assess the projected consequences of candidate control actions. Offline evaluation on held-out data collected in the CARLA simulator compares three group-based RL methods. The results show improvements in both action prediction accuracy and encoded-rule compliance over the supervised baseline. The best-performing variant increases the overall rule pass rate from 92.9% to 97.6%, with zero observed violations of the encoded road-user-priority check in the evaluated samples. These findings support symbolic rule feedback as a means of improving VLA policy alignment with explicit traffic-rule criteria.
Figures & tables
Fig. 1 : Overview of the proposed rule-guided VLA framework. SFT (top) trains the model to predict driving actions and generate reasoning outputs from visual observations and textual inputs. RLFT (bottom) refines the policy using symbolic traffic-rule feedback and an imitation penalty that encourages consistency with expert actions.
Fig. 2 : An illustrative example of translating a UK Highway Code rule into symbolic predicates, contextual factors, and an executable compliance check.
Fig. 3 : An illustrative example of the trajectory predictions made by the SFT, GRPO, GDPO and Dr. GRPO pretrained VLA models. The figures show the throttle, steering, projected speed and the trajectory waypoints.
Fig. 4 : Training dynamics of the proposed rule-compliant VLA method. (a) SFT loss trajectory. (b) Reward curves for the subsequent RLFT stage.
Fig. 5 : Per-rule violation rates over 15 RLFT epochs. Rates are computed over all symbolic rollout evaluations where the rule was active.
Fig. 6 : Distribution of longitudinal and lateral driving decisions learned by each RL optimization method. Longitudinal actions are grouped into braking, coasting, and acceleration. Lateral actions are grouped into left, straight, and right steering.
Metric
SFT
GRPO
GDPO
Dr. GRPO
Throttle MAE (%)
47.38±0.00
21.19±0.05
20.77±0.02
19.16±0.02
Steer MAE (%)
18.70±0.02
9.17±0.10
7.11±0.10
5.36±0.20
Throttle RMSE (%)
48.31±0.05
23.71±0.08
23.54±0.05
23.17±0.00
Steer RMSE (%)
20.96±0.10
13.74±0.10
12.95±0.10
12.74±0.10
Euclidean MAE (%)
73.39±0.05
43.07±0.05
42.30±0.05
40.57±0.05
Euclidean RMSE (%)
81.07±0.05
45.82±0.20
45.43±0.20
45.01±0.10
TABLE I : VLA action regression error and rule pass/violation rates for SFT, GRPO, GDPO, and Dr. GRPO, reported for about 3,600 held-out samples. To estimate training variance, each method was trained with three independent random seeds controlling dataset ordering and action noise sampling. All runs were initialized from the same supervised fine-tuned checkpoint. Values are reported as mean ± standard deviation.
Rule
SFT
GRPO
GDPO
Dr. GRPO
Safe Speed Rule Score
0.9200
1.0000
1.0000
1.0000
Safe Distance Rule Score
0.7468
0.7848
0.7975
0.7975
Hierarchy of Road Users Rule Score
-1.0000
0.8820
0.8820
1.0000
TABLE II : Symbolic compliance score (Eq. ( 4 )) for each UK Highway Code rule, averaged over frames in which the rule is active, for SFT, GRPO, GDPO, and Dr. GRPO. Higher is better; +1.0 indicates satisfaction in every active frame.
Control Accuracy
Rule Violation Rate ↓
Rules in Reward
Throttle MAE (%) ↓
Steer MAE (%) ↓
Rule H
Rule 125
Rule 126
Rule-H + Rule-125 + Rule-126
19.29±0.02
6.50±0.02
0.00±0.10%
0.10±0.10%
5.90±0.30%
Rule-125 + Rule-126
21.04±0.02
5.67±0.01
6.10±0.30%
0.20±0.10%
5.50±0.20%
Rule-H + Rule-126
21.54±0.02
6.75±0.02
5.40±0.20%
0.20±0.10%
5.70±0.20%
Rule-H + Rule-125
20.61±0.02
6.98±0.02
4.40±0.20%
0.10±0.10%
8.00±0.40%
TABLE III : Rule ablation study for Dr. GRPO. Each row corresponds to a variant trained with a different subset of rules in the reward signal. Violation rates and control accuracy (MAE) are reported on the held-out test set. Values are reported as mean ± standard deviation over three random seeds. ↓ denotes lower is better.
Fig. 7 : Per-condition rule compliance for the SFT checkpoint and the change ( Δ ) introduced by each RL variant (GRPO, GDPO, Dr. GRPO), grouped by UK Highway Code rule. The leftmost column reports the absolute SFT pass rate (%); the remaining columns report the pass-rate change relative to SFT in percentage points.
Autonomous driving develops vehicles that perceive road scenes, predict interactions, and plan and control safe movement with limited or no human input.
Vision-Language-Action (VLA) models have emerged as a promising paradigm for end-to-end autonomous driving. However, existing VLAs' training relies heavily on text-centric visual question answering and chain-of-thought reasoning data, which emphasizes linguistic reasoning rather than action-grounded planning. As a result, the learned representations capture semantic knowledge but lack spatial dependencies crucial for reliable trajectory prediction. We propose DriveTeach-VLA, a framework that explicitly teaches VLAs what to see and where to look. Driving-aware Vision Distillation (DVD) injects driving-specific perceptual priors into the vision encoder, while 2D Trajectory-Guided Prompts (2D-TGP) provide spatial conditioning aligned with feasible driving trajectories. Together, they form a vision-guided learning pipeline: what to see (DVD pretraining) - where to look (TGP-guided SFT) - how to act (TGP-guided GRPO). DriveTeach-VLA achieves the state-of-the-art performance on NAVSIM and nuScenes. Our code is available at: https://github.com/ShivaTeam/DriveTeach-VLA.
Yuguang Yang, Canyu Chen, Zhewen Tan +10
School of Electronic Information Engineering, Beihang University · Institute for AI Industry Research (AIR), Tsinghua University · National College for Excellent Engineers, Beihang University +5
Driving Vision-Language-Action Models (Driving VLAs) aim to use language to improve end-to-end planning, but the language-action gap limits this promise. We propose DriveMA, a Driving VLA framework built on verifiable meta-actions, which summarize future ego motion into compact language-domain intentions and can be constructed from expert trajectories with a trajectory-grounded annotation pipeline and can be verified against generated trajectories through rule-based projection. DriveMA exploits this verifiability with action-centric supervised training and a data-efficient turn-level credit assignment reinforcement learning framework, explicitly aligning high-level decisions with low-level trajectory planning through dense rewards and precise credit assignment. DriveMA sets a new state of the art on the Waymo Open Dataset Vision-based E2E Driving, achieving a Rater Feedback Score of 8.060 with a 2B model and further improving it to 8.079 with a 4B model; it also obtains competitive closed-loop planning performance on NAVSIM. These results show that even a simple meta-action interface can achieve state-of-the-art planning when made verifiable and optimized for language-action alignment. Code, data, and models are available at https://tsinghua-mars-lab.github.io/DriveMA.
Weicheng Zheng, Yixin Huang, Qiao Sun +2
1Shanghai Qi Zhi Institute · 3Tongji University · 2IIIS Tsinghua University
Vision-language-action (VLA) models generate chain-of-thought (CoT) reasoning alongside driving trajectories, but existing benchmarks evaluate only trajectory quality and do not assess whether the CoT is relevant, consistent, or causally connected to the driving action. We introduce VLADriveBench, a framework that combines observational metrics (mentioning, hallucination, contradiction, action alignment) with a CoT intervention protocol to provide complementary views of the CoT-action relationship. Applying VLADriveBench to three models across two architectures, we find that the two analyses can diverge sharply: ORION scores highest on observational alignment yet its CoT is epiphenomenal, while Alpamayo v1.5 scores lower yet its CoT is strongly causal, with visual salience gating the extent of CoT influence.