A robot policy with controllable speed — "TempoVLA," a speed-controllable Vision-Language-Action model

This work (TempoVLA) lets a robot policy vary its motion speed by situation.

Robot manipulation alternates between low-risk transit phases that call for fast execution and high-risk contact stages that demand slow, precise motion. Yet existing Vision-Language-Action models (VLAs) only inherit a single fixed speed from training demonstrations. Prior efforts to accelerate VLAs (model compression, KV-cache reuse, reinforcement learning) only shift the policy from one fixed speed to another, leaving deceleration almost unexplored.

The authors observe that the magnitude of each predicted action already governs how fast the robot moves, opening a direct route to controllable execution speed. TempoVLA, a single VLA whose speed is set by an explicit condition, combines two coupled components: (1) a data-side Variable-Speed Trajectory Augmentation (VSTA) that re-times demonstrations to any target speed by merging or splitting actions while preserving motion semantics; and (2) a model-side conditioning mechanism that feeds the speed to the policy. Statistics show VSTA reaches the requested speed with negligible motion error.

Experiments in simulation and on real-world tasks show TempoVLA achieves flexible speed control in both directions, while VSTA also boosts default (1×) performance via better data utilization. Cooperating with a large multimodal model, TempoVLA realizes dynamic speed control — accelerating through low-risk phases and decelerating for high-risk ones.