Vision-Language-Action (VLA) models are advancing autonomous driving by replacing modular pipelines with unified end-to-end architectures. However, current VLAs face two expensive requirements: (1) massive dataset collection, and (2) dense reasoning annotations. In this work, we address both challenges with NORD (No Reasoning for Driving). Compared to existing VLAs, NORD achieves competitive performance while being fine-tuned on <60% of the data and no reasoning annotations, resulting in 3x fewer tokens. We identify that standard Group Relative Policy Optimization (GRPO) fails to yield significant improvements when applied to policies trained on such small, reasoning-free datasets. We show that this limitation stems from difficulty bias, which disproportionately penalizes reward signals from scenarios that produce high-variance rollouts within GRPO. NORD overcomes this by incorporating Dr. GRPO, a recent algorithm designed to mitigate difficulty bias in LLMs. As a result, NORD achieves competitive performance on Waymo and NAVSIM with a fraction of the training data and no reasoning overhead, enabling more efficient autonomous systems. Website: https://nord-vla-ai.github.io/
Figures & tables
Figure 1: Comparison of VLA training pipelines. (a) Existing approaches depend on large-scale reasoning data generation, followed by extensive SFT and RL fine-tuning. (b) In contrast, NoRD directly utilizes a small-scale driving dataset for SFT, and performs RL fine-tuning tailored for weak SFT policy, enabling data-efficient learning without reasoning supervision.
Figure 2: Reward distribution in the weak SFT model. The group-mean PDM score is shown with band representing the mean of the corresponding group standard deviation for NoRD-base . GRPO struggles to optimize high-variance regions (the majority) and is effective only in low-variance regions (the trajectories in green and red are for ground truth and NoRD-base prediction).
Figure 3: Evolution of group-mean PDM score during RL fine-tuning. (a) GRPO struggles to optimize samples with high group variance during training, particularly in the range [0.2 – 0.65] . (b) Dr. GRPO effectively optimizes high-variance samples during training, resulting in significant overall performance gains.
Figure 4: Qualitative comparison of RL fine-tuning (RLFT) on the weak SFT model using GRPO and Dr. GRPO. With Dr. GRPO, NoRD successfully learns complex maneuvers such as sharp turns and lane changes without collisions, whereas GRPO fails to optimize the weak SFT model ( NoRD-base ) and collides (in red).
Model
PDMS ↑
NoRD-base
76.66
NoRD-base + GRPO
77.18 (+0.67%)
NoRD-base + Dr. GRPO
85.62 (+11.68%)
Table 1: Comparison of RL fine-tuning (RLFT) on NoRD-base with GRPO and Dr. GRPO on NAVSIM test set. While GRPO fails to improve NoRD-base , we get significant gains with Dr. GRPO.
Figure 5: Model architecture of NoRD . NoRD directly predicts action tokens without requiring reasoning traces, enabling a significantly more efficient training and inference pipeline.
Model
w/o Reason
w/o Ensemble
RFS ↑
ADE@3 ↓
Poutine [ 36 ]
✗
✓
7.986
1.2055
HMVLM [ 40 ]
✗
✓
7.736
1.3269
DiffusionLTF
✓
✗
7.717
1.3561
UniPlan
✓
✗
7.692
1.3083
AutoVLA [ 51 ]
✗
✓
7.556
1.3507
NoRD
✓
✓
7.709
1.2504
Table 2: Test results on the Waymo Vision-based End-to-End Driving Benchmark. NoRD achieves competitive performance, without reasoning or ensembling.
Method
w/o R
w/o L
C
PDMS ↑
Collision ↑
DAC ↑
Direction ↑
Progress ↑
TTC ↑
Comfort ↑
BEV-based Methods
UniAD [ 17 ]
✓
✓
32
83.4
97.7
91.9
-
78.8
92.9
100
Transfuser [ 7 ]
✓
✗
3
84.0
97.7
92.8
97.9
79.2
92.8
100
Hydra-MDP [ 24 ]
✓
✗
3
86.5
98.2
96.2
95.8
78.7
94.6
100
DiffusionDrive [ 25 ]
✓
✗
3
88.1
98.2
96.2
-
82.2
94.7
88.1
VLA-based Methods
Table 3: Test results on NAVSIM benchmark (navtest subset). NoRD achieves competitive performance ( w/o R : Without reasoning data, w/o L : without LiDAR data, and C : Number of RGB frames; * BoN refers to the average over best score per sample out of 6 outputs with different random seeds).
Figure 6: Pareto-optimal curves on two driving benchmarks. (a) NoRD is the only VLA in NAVSIM operating in the high-performance, high–data-efficiency region using only RGB inputs. (b) NoRD achieves competitive RFS on WaymoE2E with a fraction of the training data, without ensembling or reasoning supervision. Shaded regions provide a qualitative categorization of model efficiency and performance for ease of visualization.
Figure 7: Qualitative Results on NAVSIM (navtest subset) . NoRD safely executes sharp turns, respects traffic lights, and avoids collisions, demonstrating robust driving behavior. The predicted trajectory is shown in red.
Figure 8: Qualitative Results on WaymoE2E test set . NoRD drives safely in challenging out-of-distribution scenarios like unsafe pedestrian crossing and construction site. The predicted trajectory is shown in red (stitched, center-cropped for visualization).
Figure 9: Comparison of token and runtime efficiency. NoRD is the most (a) token and (b) runtime efficient VLA.
Figure 10: Training improvement patterns for GRPO (top, red) and Dr. GRPO (bottom, blue) across intra-group variance levels. The y=x line indicates no change in PDM score. GRPO shows strong improvements for low-variance samples with initial scores in [0.8,1.0] (panel (a)), while Dr. GRPO outperforms GRPO for medium- and high-variance samples (panels (b) and (c)), with denser concentration above y=x .
Figure 11: Training and validation curves for RL fine-tuning with GRPO and Dr.GRPO. Dr.GRPO (in red) consistently outperforms GRPO (in blue) on the (a) training and (b) validation sets by a significant margin.
Method
PDMS ↑
Collision ↑
DAC ↑
Direction ↑
Progress ↑
TTC ↑
Comfort ↑
NoRD-base
76.66
96.45
86.37
94.62
71.58
90.37
99.97
NoRD-base +GRPO
77.18
91.89
90.12
91.84
80.06
80.13
99.96
NoRD-base +Dr. GRPO
85.62
97.56
94.92
95.94
79.30
93.53
100
Table 4: Detailed comparison of RL-fine-tuning of NoRD-base with GRPO and Dr. GRPO. Dr. GRPO based RL fine-tuning is almost always better than GRPO.
Figure 12: Example of NoRD inference. Given multi-view images, past trajectory, and the current velocity, acceleration, and driving command, NoRD directly predicts the trajectory tokens without explicit reasoning.
Vocabulary Size
PDMS ↑
512
83.07
2048
85.62
Table 5: Effect of k-disc vocabulary size on the performance of NoRD on navtest.
Metric Name
Value ↑
Construction Score
8.072616
Intersection Score
7.9252014
Pedestrian Score
7.7775736
Cyclist Score
7.8055406
Multi Lane Maneuver Score
7.8262477
Single Lane Maneuver Score
8.308635
Table 6: Detailed results on WaymoE2E Test Set.
Figure 13: Failure cases of NoRD . The predicted trajectory is shown in red and the violations marked in red circle.
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
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.
Vision-language-action models (VLA) promise human-steerable autonomous driving, but their training is bottlenecked by the scarcity of frames paired with natural-language instructions: while camera streams and expert trajectories are logged at scale, language annotations (e.g., turn left at the intersection) remain scarce and expensive to acquire. To address this challenge, we introduce Latent Action Driving Annotations (LADA), a three-stage pipeline that transforms abundant unlabelled observation-trajectory pairs into a substrate for language-conditioned control. First, we train a latent action model with a vector-quantised bottleneck, producing a compact codebook of high-level vehicle intents. Second, a small language-annotated subset is used to train a vision-language translator to map observations and language instructions into this codebook. Third, we train a driving VLA on observation-latent-action pairs over the full unlabelled corpus. Using fewer than 5% of language annotations and without leveraging any auxiliary chain-of-thought reasoning or visual question answering streams, LADA achieves a Driving Score of 87.98 and a Success Rate of 70.46% on the closed-loop Bench2Drive benchmark, matching or surpassing fully supervised baselines.
Alexey Zakharov, Kemal Oksuz, Puneet K. Dokania
Robert Bosch GmbH, Germany · Five AI Ltd., United Kingdom