SPAR: Structural Pruning and Action Recovery for Vision-Language-Action Models

Key points

• The limited onboard computing resources of robots make the inference demands of vision-language-action (VLA) models costly to accommodate.1
• Reducing these demands through aggressive compression, however, can degrade manipulation performance beyond what supervised fine-tuning alone recovers.1
• The compressed policy achieves a $1.98\times$ full-policy inference speedup over the unpruned model.1
• These results identify substantial intermediate-dimension redundancy and show that post-pruning RL can recover manipulation performance in compact policies.1

Sources (1)

• [1]SPAR: Structural Pruning and Action Recovery for Vision-Language-Action Models
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 02:28 PM

  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.

  The limited onboard computing resources of robots make the inference demands of vision-language-action (VLA) models costly to accommodate.

Extractive summary: sentences quoted from the sources.