Reproduce “Learning Continuous Control Policies by Stochastic Value Gradients” 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

Nicolas Heess, Greg Wayne, David Silver, Timothy Lillicrap, Yuval Tassa, Tom Erez

Stochastic Value Gradients (SVG) train stochastic continuous-control policies by expressing randomness as explicit noise and differentiating a learned dynamics model around observed transitions. SVG(1) uses a value function for rewards beyond one modeled step, rather than relying on model-generated trajectories. In simulation, it handles reduced model capacity better than SVG(∞), but these results do not establish physical-robot performance or robustness to arbitrary model errors.

Summary by alphaXiv.

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