Distributed Reinforcement Learning for Robotic Assembly
summary
Three projects show concrete results: Project Touch trains using Apex DDPG plus preserved human demos and deploys policies that succeed on 75° and 60° tight lap joints but fail a 45° case; adding robot-collected successes yields dynamic experience replay, boosting packing-hole success from 58% to 91% and improving several buffer schemes; adding recurrence and a two-level prioritized buffer yields rd2, which is the only method to solve an 8° angular-offset peg task and retains ~100% success on 5 mm and 10 mm offsets and ~90% under 50% torque noise.
AI-written summary · about
chapters (13)
- 0:00Intro
- 1:25Autodesk Al Research Focus Areas
- 2:26Project Touch
- 4:19Ape-X DDPG with Human Demonstrations
- 6:02Challenge: Misleading human demonstration
- 6:52Dynamic Experience Replay (DER) Robots learn from each other's successful experiences
- 7:28Four Different Buffer Structures
- 8:01System Diagram built on top of Ray's implementation of Ape-X DDPG
- 8:58Results: Peg-in-Hole
- 11:27Robotic Assembly with Partial Observability
- 14:29Training RD2 Agents
- 16:30Evaluations
- 19:51Recap
related talks
14mHuman-in-the-Loop Reinforcement Learning
18mRobotics: why now?
15mRobot Hand Driven by Computer Vision and Reinforcement Learning
16mReal-World Robotics as the Next Frontier for AI?
17mBeyond Automation: AI and the Evolution of Robotic Intelligence
16mKeynote: Robot Foundation Models
21mProgramming-by-Demonstration for Long-Horizon Robot Tasks
28mPreference Learning from Minimal Human Feedback for Interactive