Overview
Field: Machine Learning / Robotics Authors: Dong Jing, Jingchen Nie, Tianqi Zhang Published: 2026-06-04 arXiv: 2606.06491
Summary
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 through model compression, KV-cache reuse, or reinforcement learning only shift the policy from one fixed speed to another, and deceleration remains 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. They turn this observation into TempoVLA, a single VLA whose execution speed is controlled by an explicit condition.
TempoVLA combines two coupled components:
1. Data-side Variable Speed Trajectory Augmentation (VSTA): retimes demonstrations to arbitrary target speeds by merging or splitting actions, while preserving motion semantics. Statistics show VSTA reaches target speeds with negligible motion error. 2. Model-side conditioning mechanism: feeds the desired speed into the policy.
Results
- Simulation and real-world experiments show TempoVLA achieves bidirectional, flexible speed control.
- VSTA also improves default 1x-speed performance through better data utilization.
- Paired with a large multimodal model, TempoVLA enables dynamic speed control: accelerating during low-risk transit phases and decelerating during high-risk contact phases.