Reproduce “Shared Autonomy via Deep Reinforcement Learning” 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
Siddharth Reddy, Anca D. Dragan, Sergey Levine
Researchers at the University of California, Berkeley developed a model-free deep reinforcement learning framework for shared autonomy, enabling an assistive agent to collaborate with human users without prior knowledge of environmental dynamics, user goals, or user behavior policies. The system demonstrated improved human task success rates and reduced crash rates in both game simulations and physical robot control tasks.
Key insights
- Deep Reinforcement Learning can effectively infer implicit user intent and adapt to unknown environmental dynamics in real-time by treating user control as an additional sensor input.
- Maintaining a user's real-time feedback loop is crucial; this can be achieved by allowing the agent to select actions that are 'close enough' to the user's suggestions while still being highly rewarded by the agent's Q-function.
- Decomposing the reward function into general, known terms and sparse human feedback allows for pretraining the agent in simulation to improve sample efficiency and reduce the human interaction burden.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.