TempoVLA: Learning Speed-Controllable Vision-Language-Action Policies

Existing Vision-Language-Action (VLA) models for robot manipulation typically operate at a single, fixed execution speed derived from training demonstrations. This limitation presents a challenge in robotic tasks, where optimal performance often necessitates dynamic speed adjustments, such as rapid transit during low-risk phases and precise, slow movements during high-risk contact stages. Prior attempts to modify VLA speeds have primarily focused on shifting to a different fixed speed, with deceleration capabilities remaining largely unexplored. This research introduces TempoVLA, a novel VLA framework designed to enable explicit and controllable execution speed. The core insight is that the inherent magnitude of predicted actions directly influences robot movement speed, thereby offering a direct pathway for

speed control. TempoVLA integrates two primary components: a data-side Variable-Speed Trajectory Augmentation (VSTA) and a model-side conditioning mechanism. VSTA functions by re-timing demonstration trajectories to a specified target speed, intelligently merging or splitting actions while rigorously preserving the original motion semantics. Concurrently, the model-side conditioning mechanism explicitly feeds the desired speed parameter to the policy, allowing for dynamic adaptation. Empirical evaluations indicate that VSTA effectively achieves the requested speed with minimal deviation in motion. Furthermore, experiments conducted in both simulated and real-world environments demonstrate that TempoVLA facilitates flexible speed control across a spectrum of velocities, encompassing both acceleration and deceleration. This approach significantly enhances the adaptability and operational safety of

robotic manipulation systems by enabling on-the-fly speed modulation based on task requirements.

Message: Hi FE, I need help with “TempoVLA: Learning Speed-Controllable Vision-Language-Action Policies” in “Electrical Engineering”