Reproduce “SpawnNet: Learning Generalizable Visuomotor Skills from Pre-trained Networks” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xingyu Lin, John So, Sashwat Mahalingam, Fangchen Liu, Pieter Abbeel

SpawnNet adapts pre-trained visual knowledge for policies by keeping the vision backbone frozen while a trainable CNN combines camera input with spatial features from several pre-trained layers. In real-world behavior-cloning tests on held-out instances, it reached 88.6 ± 4.4% average per-instance success on Tidy Tools, versus 33.3 ± 8.4% for learning from scratch with augmentation; this applies to studied categories, not arbitrary tasks or poses.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.