TempoVLA: A Speed-Controllable Vision-Language-Action Policy for Robot Manipulation

Forum topic · 小凯 · 2026-06-06

Summary

TempoVLA (arXiv:2506.08295) is a vision-language-action (VLA) model that lets robots explicitly control their execution speed during manipulation tasks. Robot manipulation naturally alternates between low-risk transit phases that benefit from fast motion and high-risk contact phases that require slow, precise movement, but existing VLA policies inherit a single fixed speed from their training demonstrations. The authors observe that the magnitude of each predicted action already governs how fast the robot moves, opening a direct path to controlling execution speed. TempoVLA combines a data-side Variable-Speed Trajectory Augmentation strategy with a model-side speed-conditioning mechanism, enabling a single policy to operate at variable speeds on demand. The work was authored by Dong Jing, Jingchen Nie, and Tianqi Zhang, published on arXiv in June 2025 in the machine learning domain.

Paper Overview

Research area: Machine Learning (ML) Authors: Dong Jing, Jingchen Nie, Tianqi Zhang Published: 2025-06-11 arXiv: 2506.08295

Introduction

Robot manipulation naturally alternates between two kinds of motion:

Low-risk transit phases — where fast execution is desirable
High-risk contact stages — where slow, precise motion is required

However, existing vision-language-action models (VLAs) only inherit a single fixed speed from their training demonstrations, giving them no way to adjust execution speed at deployment time.

Key Idea

The authors observe that the magnitude of each predicted action already controls how fast the robot moves. This opens a direct path to explicitly controlling execution speed within a VLA policy.

TempoVLA

TempoVLA is a single VLA whose execution speed is controlled by an explicit condition. It combines two components:

1. Variable-Speed Trajectory Augmentation (data side) — augmenting training data with trajectories at varying speeds 2. Speed-conditioning mechanism (model side) — conditioning the policy on a desired execution speed

Original 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 VLAs only inherit a single fixed speed from training demonstrations. We propose TempoVLA, a single VLA whose execution speed is controlled by an explicit condition, combining Variable-Speed Trajectory Augmentation and a model-side conditioning mechanism.