Reproduce “AI Coaching for Accelerating Human Skill Development with 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
Wei Wang, Enlin Gu, Antonio Loquercio, Haimin Hu, Rahul Mangharam
Researchers from the University of Pennsylvania developed an AI coaching framework, Learning to Coach (L2C), which formalizes AI-human interaction as a non-cooperative dynamic game to prioritize human skill development. When applied to FPV drone racing, L2C coaching reduced human learner lap times by 27.9% and failure counts by 3.52 per lap, outperforming existing shared-control assistance methods.
Key insights
- Effective AI coaching necessitates a strategic balance of 'scaffolding' (providing assistance) and 'stepping back' (allowing productive failures), dynamically adapted to the learner's evolving skill.
- A non-cooperative game formulation, with distinct objectives for the AI (long-term competence) and the human (immediate performance), is critical for the AI to prioritize pedagogical goals.
- Coaching policies can be tractably trained using model-free RL by reformulating the problem into a single-agent POMDP and utilizing a surrogate reward function (skill level difference) that formally correlates with the Value of Independence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.