Reproduce “1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities” 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
Kevin Wang, Ishaan Javali, Michał Bortkiewicz, Tomasz Trzciński, Benjamin Eysenbach
The authors scale CRL for online goal-conditioned learning in simulated tasks. Agents learn from their own trajectories, contrasting state-action pairs and future goals with mismatched pairs. On Humanoid, scaling actor and critics from depth 4 to 64 raises Time at Goal from 12.6 ± 1.3 to 649 ± 19, a 52× improvement; offline CRL shows little evidence of depth gains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.