FastOPD: On-Policy Distillation for Lightweight VLA Deployment

Key points

• Vision-Language-Action (VLA) foundation models have scaled rapidly to enhance manipulation performance and generalizability, but this scaling incurs high computational costs that render real-world deployment increasingly challenging.1
• Specifically, FastOPD adapts a flow map for single-state teacher supervision and combines it with a self-consistency objective to construct a compact student that learns the teacher dynamics.1
• We evaluate FastOPD across diverse foundation policies in simulation and real-world experiments.1
• We further demonstrate its applicability to a World Action Model (WAM) and deploy a compact student distilled from MolmoAct2 on a real robot.1

Sources (1)

• [1]FastOPD: On-Policy Distillation for Lightweight VLA Deployment
  Hugging Face Daily Papers · Oct 2, 12:00 AM

  In this work, we propose FastOPD, a foundation-to-lightweight VLA framework that enables the practical deployment of large-scale VLAs through efficient on-policy distillation.

  Vision-Language-Action (VLA) foundation models have scaled rapidly to enhance manipulation performance and generalizability, but this scaling incurs high computational costs that render real-world deployment increasingly challenging.

Extractive summary: sentences quoted from the sources.