Existing Vision-Language-Action (VLA) models inherit a single fixed execution speed from their training data. We propose TempoVLA, a single speed-controllable VLA framework that pairs a data-side Variable-Speed Trajectory Augmentation (VSTA) with a lightweight model-side conditioning mechanism, both lightweight and applicable to existing VLAs. Experiments in simulation and real world show that TempoVLA delivers flexible bidirectional speed control ability, improves default baseline performance, and achieves dynamic speed control with an external VLM.
Introduction and Motivation
Vision-Language-Action models trained on offline robot demonstration data learn a single fixed execution speed from the data distribution. Once deployed, these policies execute actions at the same pace regardless of task phase or environmental conditions, limiting their flexibility and robustness. TempoVLA addresses this limitation by introducing speed controllability directly into the VLA framework, enabling the robot to vary its execution speed on command during deployment.
TempoVLA Framework
The core insight behind TempoVLA is that speed control should operate at two levels: augmenting the training data to include variable-speed trajectories, and conditioning the policy outputs on a desired speed command. The Variable-Speed Trajectory Augmentation (VSTA) component re-times existing demonstration data at multiple speeds by accumulating and splitting end-effector (EEF) actions. The model-side conditioning mechanism takes a scalar speed factor as input and modulates the policy's action predictions accordingly.
Why End-Effector Space for Speed Augmentation
For manipulation tasks, task success depends on the relative pose between the end effector and manipulated objects. Re-timing demonstrations in end-effector space rather than joint space preserves task semantics more faithfully. Three structural reasons explain this preference:
- Geometric amplification at single joints: Small re-timing errors on proximal joints are geometrically magnified at the end effector due to lever arm effects.
- Non-linear coupling across the kinematic chain: Forward kinematics maps joint angles non-linearly to end-effector pose, so linearly interpolated joint increments produce non-linear end-effector motion.
- Controller realizability: End-effector-space re-timing lets the low-level controller use inverse kinematics to find dynamically feasible joint trajectories at the new speed.
While VSTA can still be applied directly on platforms that expose only joint commands, end-effector actions produce better task-relevant performance.
Failure Mode Analysis
Analysis of qualitative failure cases reveals the usable speed envelope for TempoVLA. At slow commands, the robot may exhibit stalled progress — each action contributes too little effective task progress, trapping the policy near a phase boundary instead of transitioning to the next manipulation stage. Slow execution is beneficial only when the reduced action magnitude still preserves enough progress to complete required phase transitions.
At fast commands, the end effector can overshoot the target region, passing the valid interaction window before the policy can correct. This is particularly damaging in contact-rich stages where successful interaction depends on narrow spatial and temporal tolerances. Once the gripper moves past the object or perturbs it into an out-of-distribution state, the remaining rollout can become unrecoverable.
Implications for Speed Selection
TempoVLA should not be used by assigning an extreme fixed speed to an entire rollout. Instead, speed should be selected according to the current manipulation phase: faster during low-risk free-space motion and slower near contact-rich phases requiring precise interaction. This observation is consistent with dynamic speed scheduling results, where phase-aware speed selection outperforms fixed-speed execution.
Relationship with Controller-Focused Methods
Controller-side approaches — scaling target velocities or stretching step periods after the policy produces outputs — sit downstream of the policy. These methods can only rescale or retime what the policy has already predicted, while variable-speed training changes the content of those predictions at the policy level. The two approaches are orthogonal and can be composed directly: a TempoVLA policy can still be paired with any controller-side modulation when finer execution-side tuning is desired.
Limitations and Future Work
At the high end of the speed range, the realized speedup gradually saturates because the policy's per-step targets exceed the fixed low-level controller's tracking bandwidth. Co-tuning the controller alongside TempoVLA is a natural extension for future work.
Frequently Asked Questions
How does TempoVLA differ from simply scaling the velocity commands after the policy produces them? TempoVLA operates at the policy level during training, changing the content of action predictions themselves. Controller-side velocity scaling acts downstream and can only rescale outputs the policy already generated, making the two approaches complementary.
Why does TempoVLA augment trajectories in end-effector space rather than joint space? End-effector space directly preserves task-relevant quantities like relative pose between gripper and objects, avoids geometric amplification from proximal joints, and lets the low-level controller dynamically resolve feasible joint trajectories at new speeds.
What happens if you command an extreme fixed speed for an entire rollout? Slow fixed speeds can cause stalled progress where each action contributes too little task advancement, while fast fixed speeds cause overshoot failures where the gripper moves past objects. Speed should be adjusted per manipulation phase.
Can TempoVLA be combined with existing Vision-Language-Action models? Yes, the VSTA data augmentation and model-side conditioning are designed as lightweight modules that apply to existing VLA frameworks without architectural overhauls.