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

TempoVLA introduces a controllable speed mechanism for Vision-Language-Action robot models, enabling flexible execution from fast transit to slow precision work. The approach uses trajectory augmentation during training and conditioning mechanisms during inference, allowing a single model to dynamically adjust operational speed based on task risk levels.

TempoVLA addresses a fundamental limitation in current robotic AI systems: the inability to dynamically adjust execution speed based on task requirements. Traditional Vision-Language-Action models inherit fixed speeds from training data, creating inefficiencies when robots must balance speed with safety. This research demonstrates that action magnitude already encodes speed information, providing a natural control lever previously unexploited.

The technical contribution spans two complementary approaches. Variable-Speed Trajectory Augmentation (VSTA) re-times training demonstrations by merging or splitting actions while preserving semantic meaning, effectively creating synthetic training data across speed ranges. The model-side conditioning mechanism then learns to interpret explicit speed instructions. This dual-component design avoids costly retraining while maintaining motion quality across different execution tempos.

The implications extend beyond robotics efficiency. Dynamic speed control enables robots to accelerate through low-risk phases like transit while automatically decelerating for high-risk contact operations. This mirrors human motor control patterns and could significantly reduce manipulation errors and equipment damage. Integration with multimodal language models enables contextual speed decisions based on task understanding rather than pre-programmed parameters.

The research validates that improved data utilization through VSTA boosts baseline performance even at default speeds, suggesting the augmentation technique provides genuine regularization benefits beyond speed control. For robotics developers and companies, this offers a path to more capable, safer robotic systems without architectural redesign. The work also hints at broader applications in autonomous systems where context-dependent execution speed matters—from drones to manufacturing equipment.

• →TempoVLA enables single Vision-Language-Action models to execute at variable speeds controlled by explicit conditioning inputs
• →Variable-Speed Trajectory Augmentation creates training data across speed ranges by merging/splitting actions while preserving motion semantics
• →Dynamic speed control allows robots to automatically accelerate through low-risk transit phases and decelerate for high-risk contact tasks
• →VSTA improves default performance through better data utilization, boosting baseline speed even without speed modulation
• →Integration with multimodal models enables context-aware speed decisions based on task understanding rather than fixed policies