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

A research team from ETH Zurich and KU Leuven has introduced AR-VLA, a novel architecture designed to give robots a crucial capability they often lack: a coherent memory of their own actions. Current Vision-Language-Action (VLA) models and diffusion policies operate reactively, resetting their temporal context with every new camera frame. This leads to jerky, inconsistent movements. AR-VLA solves this by functioning as a standalone autoregressive 'Action Expert' that generates actions as a continuous causal sequence, maintaining a long-lived internal history. This allows it to produce smoother, more context-aware motion trajectories.

The core innovation is a decoupled design that addresses the frequency mismatch between fast control (needed for movement) and slow reasoning (for task understanding). The Action Expert handles the 'kinematic syntax' of movement independently, and can be integrated with heavy perception backbones like VLMs. A key technical challenge was synchronizing these asynchronous vision, language, and action streams; the team solved this with a 're-anchoring' mechanism that mathematically accounts for perception staleness. In experiments on simulated and real-robot manipulation tasks, AR-VLA successfully replaced traditional chunk-based action heads, delivering up to 40% smoother movements while maintaining or exceeding the task success rates of top reactive models.

This work provides a robust, scalable structural foundation for the next generation of robotic policies. By ensuring spatio-temporal consistency, it moves robots closer to performing complex, long-horizon tasks with the fluidity and foresight of a human, making them more reliable and effective in dynamic real-world environments.