PaceVLA: Adaptive Temporal Action Quantization in Vision-Language-Action Models
A KAIST and RLWRLD preprint lets a robot policy decide, step by step, when it can move in coarse fast strides and when it must slow down, raising an ALLEX humanoid's conveyor throughput to 6.3 packages a minute from 3.6 while success rose to 94% from 86%. On a Franka arm it cut an air-fryer task to 14.6 seconds from 24.6 with no loss in success.
Can a vision-language-action (VLA) policy run faster than the human demonstrations it learned from without losing accuracy, by choosing the time granularity and look-ahead of its actions separately for each phase of a task?
VLA policies copy teleoperated demonstrations, and human operators move slowly and cautiously through limited-rate interfaces. The trained robot therefore executes many small steps even when it is just reaching or carrying in free space. Merging consecutive actions into fewer, larger commands speeds things up, but doing it everywhere throws away the fine motion needed at grasp, placement and contact.
Fixed temporal action quantization merges actions at one ratio for the whole trajectory. DemoSpeedup speeds up the training demonstrations using the action entropy of a proxy policy, and SAIL switches between preset slow and fast speeds using motion and gripper cues. All of them decide the speed from statistics of the actions rather than from what the robot sees in the scene.
PaceVLA fine-tunes an existing flow-matching VLA (NVIDIA's GR00T N1.5, and RLWRLD's RLDX-1-IMG in simulation) and gives its shared action expert four output decoders: two fine-grained and two quantized modes, each with a short and a long horizon. A quantization judge, trained on labels produced by a large vision-language model, decides from the camera images and the instruction whether coarse execution is safe right now, and a router then picks one mode at every inference step. Reaching and carrying get coarse, longer predictions; grasping and placing get fine, shorter ones with more frequent re-planning.
On the ALLEX humanoid with 15-DoF hands, handling 10 packages per trial, throughput rose to 6.3 packages per minute from 3.6 for the GR00T N1.5 base (1.75x) and success to 94% from 86%; fixed quantization reached 4.2 a minute at 72% and SAIL 4.3 at 78%. On a Franka arm (32 trials per task) execution time fell 37% on pick-and-place (12.1 s to 7.6 s), 41% on the air fryer (24.6 s to 14.6 s) and 45% on cup stacking, with success of 84.4%, 81.3% and 81.3% against 81.3%, 81.3% and 84.4% for the base; fixed quantization lost 28.2 and 21.9 points on the air fryer and cup tasks. On a bimanual OpenArm with 6-DoF hands time fell 33% and 39%. In simulation the effective speedup on RoboCasa pick-and-place was 1.16 with success 55.0% against 52.3%, and 1.11 on the DexJoCo dexterous suite. Per-step latency rose to 42.15 ms from 40.07 ms on an RTX 5090.
This is a preprint. The real-robot tests cover six tasks with 32 trials each and a single humanoid task, all on the authors' own setups and demonstrations (320 for the conveyor). The judge depends on labels written by a large vision-language model, so its quality hinges on that annotation pipeline. In simulation PaceVLA trails the base model in some cells, for example 45.0% against 48.5% on RoboCasa pick-and-place with RLDX-1-IMG, and the reported speedups are relative to slow teleoperated demonstrations, not to hand-tuned industrial motion programs.
Cycle time is the number that decides whether a learned policy can replace a programmed robot on a line, and this work attacks it without new hardware or new data. It matters directly to RLWRLD, whose RLDX models the paper uses, and to anyone shipping GR00T-based policies, since the method is a fine-tuning add-on with about 5% inference overhead. For Korean humanoid makers such as WIRobotics, which MEKAWIRE's profile lists as the developer of ALLEX, a near doubling of logistics throughput on the same robot is the kind of gain that moves pilots toward paid deployments within one to two years.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.