TempoBridge: Language-Guided Tempo Control for Vision-Language-Action Policies

Yeonseo Lee; Hyosup Shin; Guebin Hwang; Sungho Jo

cs.ROOct 7, 2026

TempoBridge: Language-Guided Tempo Control for Vision-Language-Action Policies

Authors: Yeonseo Lee, Hyosup Shin, Guebin Hwang, Sungho Jo

Organizations: Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea.

Abstract

Vision-Language-Action (VLA) models are effective at understanding what task to perform, but provide limited control over how it should be executed, such as moving quickly or slowly. We introduce TempoBridge, a lightweight framework that uses frozen VLA representations to modulate actions according to tempo cues in the instruction at each task phase, without additional tempo-conditioned robot demonstrations or tempo-specific base-policy fine-tuning. TempoBridge extracts tempo cues from contextual VLM representations, aligns them with task progress through a causal phase router, and modulates nominal motion commands during execution. Across LIBERO tasks, TempoBridge improves Tempo Success Rate from 52.6% to 89.7% under canonical tempo instructions while retaining high task success. It also preserves near-baseline performance when no tempo cue is present and generalizes to unseen tempo expressions without additional training. Experiments on a physical robot further demonstrate language-conditioned tempo modulation in real-world manipulation.

Figures & tables

Fig. 2: Overview of TempoBridge. Using the frozen VLM backbone within $\pi_{0.5}$ , a prototype-based tempo readout extracts an ordered tempo sequence from the instruction once and caches it for execution. At each replan, a causal phase router uses visual features reused from the policy forward pass and robot state to track task progress and select the active event. The corresponding tempo coefficient scales the translational components of the nominal action chunk.

Method	Spatial		Object		Goal		Long		Overall
Method	Task Success Rate $\uparrow$	Tempo Success Rate $\uparrow$	Task Success Rate $\uparrow$	Tempo Success Rate $\uparrow$	Task Success Rate $\uparrow$	Tempo Success Rate $\uparrow$	Task Success Rate $\uparrow$	Tempo Success Rate $\uparrow$	Task Success Rate $\uparrow$	Tempo Success Rate $\uparrow$	Task & Tempo Success Rate $\uparrow$
$\pi_{0.5}$	94.0	48.9	96.0	49.5	93.0	57.8	91.5	54.4	93.6	52.6	48.8
TempoBridge	94.5	91.0	97.5	90.8	92.5	93.8	87.0	82.9	92.9	89.7	82.0

TABLE I: Task, tempo, and task & tempo success rates (%) under quickly / slowly instructions.

Fig. 3: Qualitative examples of TempoBridge execution on LIBERO. The left two panels use quickly / slowly cues, while the right three use tempo expressions unseen during readout construction. Each panel shows the instruction and executed TCP trajectory, colored by measured TCP speed from slower motion (blue) to faster motion (red).

Method	Spatial	Object	Goal	Long	Overall
$\pi_{0.5}$	95.0	97.0	95.0	93.0	95.0
TempoBridge	98.0	96.0	95.0	88.0	94.3

TABLE II: Task success rates (%) on LIBERO benchmark instructions without tempo cues.

Method	Spatial		Object		Goal		Long		Overall
Method	Task Success Rate $\uparrow$	Tempo Success Rate $\uparrow$	Task Success Rate $\uparrow$	Tempo Success Rate $\uparrow$	Task Success Rate $\uparrow$	Tempo Success Rate $\uparrow$	Task Success Rate $\uparrow$	Tempo Success Rate $\uparrow$	Task Success Rate $\uparrow$	Tempo Success Rate $\uparrow$	Task & Tempo Success Rate $\uparrow$
$\pi_{0.5}$	95.6	50.3	99.3	49.5	96.1	61.2	89.5	53.1	95.1	53.5	50.6
TempoBridge	94.8	79.7	96.1	75.2	93.8	91.0	84.8	74.2	92.3	80.1	72.9

TABLE III: Task, tempo, and task & tempo success rates (%) for tempo expressions unseen during readout construction.

Tempo expression	Target tempo	Accuracy $\uparrow$
rapidly	FAST	98.3
swiftly	FAST	98.3
carefully	SLOW	57.9
at a slower pace	SLOW	95.0

TABLE IV: Cue-specific tempo recognition accuracy (%) for expressions unseen during readout construction.

Setting	Readout Acc. $\uparrow$	Balanced Phase Agreement $\uparrow$	Median Speed Ratio
Canonical	97.5%	89.7%	1.58 $\times$
Unseen	78.4%	90.4%	1.52 $\times$

TABLE V: Component-wise diagnostics of TempoBridge under canonical and unseen tempo expressions.

Fig. 4: Real-world experimental setup with an xArm6 robot, external and wrist-mounted Azure Kinect cameras, and the three manipulation task configurations.

Fig. 5: Real-world TCP speed under Fast

\rightarrow

Slow (FS) and Slow

\rightarrow

Fast (SF) instructions. Columns correspond to the three tasks, with the task-adapted

\pi_{0.5}

baseline in the top row and TempoBridge in the bottom row. Thin lines represent individual task-successful rollouts, connecting the mean TCP speeds of Event 1 and Event 2. Bold lines indicate the median, and error bars show the interquartile range over 10 successful rollouts per condition.

Explore similar work

TempoVLA: Learning Speed-Controllable Vision-Language-Action Policies
Jun 4, 2026Dong Jing, Jingchen Nie, Tianqi Zhang +4
1RUC · 2FDU · 3UNC
Reducing Temporal Redundancy for Efficient Vision-Language-Action Inference
Jul 14, 2026Yuzhou Wu, Yuxin Zheng, Muchun Niu +6Diffusion-Based Vision-Language-Actions Vision-Language-Action Framework
1Tianji KernalMind co ltd · 2Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · 3Shanghai Jiao Tong University, Shanghai, China +2
Latent Bridge: Feature Delta Prediction for Efficient Dual-System Vision-Language-Action Model Inference
May 4, 2026Yudong Liu, Yuan Li, Zijia Tang +12Robotic Manipulation Latent Variable
1Duke University · 2Qualcomm AI Research · University of Florida

Jun 4, 2026, cs.ROJ/K move · Enter open · S save

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

Dong Jing, Jingchen Nie, Tianqi Zhang, Jiaqi Liu+3

1RUC · 2FDU · 3UNC

Open paper AskPDF arXiv

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) A data-side Variable-Speed Trajectory Augmentation (VSTA) that re-times demonstration to any target speed by merging or splitting actions while preserving its motion semantics. (2) A model-side conditioning mechanism that feeds the speed to the policy. Statistics show that VSTA reaches the requested speed with negligible motion error. Experiments in simulation and on real-world tasks demonstrate that TempoVLA achieves flexible speed control in both directions, while VSTA additionally boosts the default

1\times

performance via better data utilization. Furthermore, by cooperating with a large multimodal model, TempoVLA realizes dynamic speed control, accelerating through low-risk phases and decelerating for high-risk ones.

Abstract

Explore similar work

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

Reducing Temporal Redundancy for Efficient Vision-Language-Action Inference

Latent Bridge: Feature Delta Prediction for Efficient Dual-System Vision-Language-Action Model Inference