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

Forum topic · 小凯 · 2026-06-07

Summary

TempoVLA (arXiv:2606.06491) is a vision-language-action (VLA) policy that gives robot manipulators explicit, bidirectional control over execution speed. Robot manipulation alternates between fast low-risk transit phases and slow, precise high-risk contact phases, but existing VLAs inherit only a single fixed speed from demonstrations; prior acceleration methods merely swap one fixed speed for another and rarely address deceleration. The authors exploit the observation that predicted action magnitude already governs movement speed. TempoVLA combines (1) Variabl[e] Speed Trajectory Augmentation (VSTA), which re-times demonstrations to arbitrary target speeds via action merging or splitting while preserving motion semantics, and (2) a conditioning mechanism that feeds a speed input into the policy. VSTA reaches target speeds with negligible motion error and improves default 1x performance through better data utilization. Combined with a large multimodal model, TempoVLA achieves dynamic speed control: acceleration in low-risk phases and deceleration in high-risk ones, validated in simulation and real-world experiments.

Paper Overview

Field: ML Authors: Dong Jing, Jingchen Nie, Tianqi Zhang arXiv: 2606.06491

Abstract

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 leave deceleration almost unexplored. We observe that the magnitude of each predicted action already governs how fast the robot moves, opening a direct route to controllable execution speed. We turn this observation into TempoVLA, a single VLA whose execution speed is controlled by an explicit condition.

TempoVLA combines two coupled components:

1. Variable Speed Trajectory Augmentation (VSTA) (data side): re-times 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. A conditioning mechanism (model side): feeds the desired speed into the policy as an explicit input.

Findings

Simulation and real-world experiments demonstrate bidirectional flexible speed control.
VSTA also improves default 1x performance through better data utilization.
Paired with a large multimodal model, TempoVLA achieves dynamic speed control — speeding up in low-risk phases and slowing down in high-risk ones.

--- *Auto-collected on 2026-06-07*