AI

FastOPD: On-Policy Distillation for Lightweight VLA Deployment

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 08/10/2026, 10:18:24 UTC.

Briefings

FastOPD is a framework for deploying large Vision-Language-Action models by using on-policy distillation to create lightweight students. The study evaluated FastOPD on LIBERO and RoboTwin 2.0, showing it retains 84% of the performance of the original model with only two inference steps. It improved the single-step success rate on RoboTwin 2.0 by 15.9 percentage points compared to the base student. The framework was tested on both simulation and real-world experiments, including a World Action Model and a real robot.

Key points

• FastOPD uses on-policy distillation to create lightweight students from large Vision-Language-Action models, retaining 84% of the original performance with only two inference steps.
• On RoboTwin 2.0, FastOPD improved the single-step success rate by 15.9 percentage points compared to the base student.

Context

Vision-Language-Action models have scaled rapidly but incur high computational costs. Existing approaches reduce costs by using smaller architectures or fewer denoising steps. FastOPD offers a new method using on-policy distillation to enable practical deployment of large-scale VLAs.

Limitations

Evaluation was conducted on simulation and real-world experiments only. Performance metrics were compared against baselines and the original model, but no long-term reliability or safety data were provided.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 07/11/2026, 11:29:54 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeiatizhqmpraa4rqb3gc2uzt34vyhqnnznykndijs44j27tdto42zy