NoRD: A Data-Efficient Vision-Language-Action Model that Drives without Reasoning

A new vision-language-action model called NoRD can match the driving performance of existing autonomous systems while using less than 60 percent of the training data and no reasoning annotations, its creators report. The model, detailed in a paper posted to arXiv, replaces the modular pipelines common in current vision-language-action (VLA) architectures with a unified end-to-end design [1]. The researchers, led by Shubh Gupta, found that standard Group Relative Policy Optimization (GRPO) failed to deliver meaningful gains when applied to policies trained on small, reasoning-free datasets [1]. They attribute that failure to difficulty bias, a phenomenon that disproportionately penalizes reward signals from scenarios that produce high-variance rollouts within GRPO [1]. To address this, NoRD incorporates Dr. GRPO, a recent algorithm designed to mitigate difficulty bias in large language models [1]. The result is a system that achieves competitive performance on the Waymo and NAVSIM benchmarks while consuming roughly three times fewer tokens than existing VLAs [1]. The paper states the model is fine-tuned on less than 60 percent of the data required by comparable systems and operates without any reasoning overhead [1]. The work was submitted to arXiv on February 24, 2026, and revised on June 5, 2026 [1].