Demystifying Reward Design in Reinforcement Learning for Upper Extremity Interaction: Practical G...
summary
Systematically trains over 450 policies across three tasks (choice reaction, pointing, tracking) and varies three reward components. Finds distance (proximity) incentives are essential to guide movement, task completion bonus is required for actual success, and an effort term is optional—it stabilizes motions and doesn’t slow completion when scaled properly. Provides sensitivity curves for component weights, shows many surprising failure modes without distance or bonus, and validates guidelines on remote-control and keyboard-typing tasks with 2M steps (vs prior 5M).
AI-written summary · about
chapters (9)
- 0:00Introduction to talk
- 0:40Motivation and problem overview
- 1:43Research questions and tasks
- 2:26Reward function components
- 3:30Results and key findings
- 6:00Practical guidelines for reward design
- 8:28Validation of guidelines
- 9:10Summary and open questions
- 10:11Q&A session
related talks
14mHuman-in-the-Loop Reinforcement Learning
06mPersonalized HeartSteps: A Reinforcement Learning Algorithm for Optimizing Physical Activity
31mOn the Complexity of Learning Good Policies With and Without Rewards
38mReinforcement Learning - What it is and How it Works
28mPreference Learning from Minimal Human Feedback for Interactive
35mIndustrial Applications of Reinforcement Learning
52mReinforcement Learning in Recommender Systems: Some Challenges
25mDesigning and Building Custom Reinforcement Learning Environments for Fine-tuning LLMs