Reproduce “Safe Deep RL in 3D Environments using Human Feedback” 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
Matthew Rahtz, Vikrant Varma, Ramana Kumar, Zachary Kenton, Shane Legg, Jan Leike
ReQueST teaches agents to avoid hazards by learning a pixel-based simulator from safe human exploration and training a reward model on sketches of imagined behavior. In a Unity apple-collection task with cliff and reward-block hazards, the authors report 3–20× fewer unsafe training events than model-free reinforcement learning. Training violations came from contractor mistakes; real-world safety remains untested.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.