Summary
TempoVLA introduces explicit execution-speed control into Vision-Language-Action models. Real manipulation alternates between fast transit phases and slow, precise contact stages, yet standard VLAs inherit a single fixed speed from demonstrations. TempoVLA conditions the policy on a target speed token, enabling the same model to adapt its pace on demand.
Key Contributions
- Variable-Speed Trajectory Augmentation (VSTA): re-times existing demonstrations to any target speed by merging or splitting action steps while preserving motion semantics, expanding training data across the speed spectrum
- Speed-conditioning mechanism that feeds an explicit speed signal into the VLA backbone so the model generates actions whose magnitude governs execution rate
- Demonstrated on multiple manipulation benchmarks, outperforming fixed-speed baselines in both task success and cycle time
Significance
First VLA to decouple task competence from execution speed, enabling adaptive pacing strategies (e.g., slow during grasp, fast during transit) from a single policy — a practical step toward deployment-ready manipulation systems.