AR-VLA: True Autoregressive Action Expert for Vision-Language-Action Models

Researchers have introduced AR-VLA, a novel standalone autoregressive Action Expert designed for Vision-Language-Action (VLA) models in robotics. Unlike existing reactive VLA models and diffusion policies that reset temporal context with each observation, AR-VLA generates actions as a continuous causal sequence using long-lived memory. This approach addresses the frequency mismatch between fast control loops and slow reasoning processes, allowing for efficient independent pretraining of kinematic syntax and modular integration with heavy perception backbones. The system employs a re-anchoring mechanism to synchronize asynchronous modalities and account for perception staleness during training and inference. Experiments on both simulated and real-robot manipulation tasks demonstrate that AR-VLA produces smoother action trajectories and superior history awareness while maintaining or exceeding the success rates of state-of-the-art reactive models. This work provides a scalable, context-aware framework for robotic policy training, effectively replacing traditional chunk-based action heads. The research highlights significant advancements in creating spatio-temporally consistent action generation, offering a robust structural foundation for future developments in generalist and specialist robotic policies.