The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Applying RL to Real-World Robotics with Abhishek Gupta - #466
Mar 22, 2021
In this discussion, Abhishek Gupta, a PhD student at UC Berkeley specializing in reinforcement learning for robotics, shares his exciting journey from Lego competitions to groundbreaking research. He dives into how robots learn reward functions from video data and the importance of supervised experts. Gupta also tackles real-world challenges of robotic learning, including multitask learning and the innovative concept of 'gradient surgery' to boost efficiency. The conversation highlights the fascinating relationship between humans and robots in everyday settings.
AI Snips
Chapters
Transcript
Episode notes
First Robotics Experience
- Abhishek Gupta's first robotics experience was in First Lego League competitions, where they hard-coded Lego robots for challenges.
- At Berkeley, he discovered more sophisticated robotic control techniques, leading him to work with Peter Abil.
RL for Real-World Robotics
- Abhishek Gupta focuses on applying RL to real-world robotics, motivated by the potential impact of robots.
- He identifies key mismatches between RL algorithm assumptions and real-world conditions as a primary research focus.
Key Research Interests
- Abhishek Gupta's research focuses on reward supervision, efficient and safe data collection, and large-scale continual learning systems.
- He explores how to bridge the gap between RL and real-world robotics by addressing these challenges.