Wang, Gu, Loquercio, Hu & Mangharam (2026) — University of Pennsylvania. cs.RO, cs.AI, cs.HC. 📄 Full text (arXiv)
This paper explores how an embodied AI agent can act as a coach that accelerates human motor-skill development using reinforcement learning. The authors argue that effective coaching requires dynamically balancing guidance with learner autonomy — too much assistance leads to over-reliance and skill atrophy, while too little leaves learners struggling.
Key findings:
- An RL-based coaching policy that adapts its level of intervention to the learner's current skill level significantly accelerates skill acquisition compared to static assistance levels.
- The AI coach that gradually fades scaffolding (consistent with scaffolding theory in intelligent-tutoring) produced the best long-term retention and transfer performance.
- Over-reliance emerged when the coach provided excessive intervention, confirming the over-reliance concern documented in generative-ai tutoring contexts.
Implications:
- RL-based coaching offers a principled framework for personalized skill development in domains beyond traditional academics (e.g., surgical training, professional-training, STEM lab skills).
- The competence-based fading policy mirrors established pedagogical best practices, suggesting formative-assessment signals can drive AI coaching adaptation.
- Extends AI safety in tutoring research to embodied/motor skill domains.
Related Pages
- scaffolding — Dynamic fading of support as competence increases
- adaptive-learning — RL-based adaptation in learning systems
- over-reliance — Risk of excessive AI assistance
- intelligent-tutoring — Core ITS principles applicable to coaching
- professional-training — Motor skill training applications
- personalized-learning — Competence-adaptive instruction
- ai-tutor-safety-harms — Over-assistance and student autonomy