Vision language action models pruned and tuned for faster robot control
SPAR: Structural Pruning and Action Recovery for Vision-Language-Action Models
RoboticsArtificial Intelligence
Summary
Robots often use large models that understand language and vision to decide what actions to take, but these models can be too big to run quickly on robots. The authors introduce a method called SPAR that makes these models much smaller by cutting out unnecessary parts while retraining them with trial and error to keep their abilities. This results in much faster robot decision-making without losing accuracy in tasks. Their approach reduces model size dramatically and improves success rates after retraining.
What this means in practice
- •For robotics engineers: Implement smaller vision-language-action models that run faster on onboard robot hardware without losing task success rates.
- •For ai system integrators: Deploy compressed language backbones and retrieve performance using reinforcement learning to speed up inference in multimodal AI systems.
Authors
Shenli Zheng, Zhou Fang, Yuan Du, Yijiang Liu, Li Du
Abstract
The limited onboard computing resources of robots make the inference demands of vision-language-action (VLA) models costly to accommodate. Reducing these demands through aggressive compression, however, can degrade manipulation performance beyond what supervised fine-tuning alone recovers. We present SPAR (Structural Pruning and Action Recovery), which simultaneously prunes the hidden and feed-forward intermediate dimensions of the large language model (LLM) backbone through pruning-aware tuning, then applies reinforcement learning (RL) to recover policy performance. Applied to $π_{0.5}$, SPAR reduces the language backbone from 2.34B to 0.09B parameters and the full model from 3.37B to 1.12B, removing approximately 96% and 67% of their parameters, respectively. At this compression level, RL raises average success across the four LIBERO suites from 84.55% to 93.45%, an 8.90-percentage-point gain. The compressed policy achieves a $1.98\times$ full-policy inference speedup over the unpruned model. These results identify substantial intermediate-dimension redundancy and show that post-pruning RL can recover manipulation performance in compact policies.