Diverse Data and Efficient Algorithms for Robot Learning

Explore diverse data collection and efficient algorithms for robot learning in this MIT Embodied Intelligence Seminar featuring Lerrel Pinto from NYU and UC Berkeley. Delve into the challenges of applying machine learning to robotics, focusing on large-scale data collection and efficient reinforcement learning for deformable object manipulation. Discover self-supervised grasping techniques, robot learning in homes, and the importance of forward models in robotic systems. Gain insights into conditional policy learning, contrastive representations, and one-step model predictive control. Examine quantitative evaluations and real-world robot experiments that demonstrate the practical applications of these cutting-edge approaches in robotics.

Intro
Progress in Robotics
Success stories in Al
Key ingredients of learning
Learning for robotics
How to grasp objects?
Self-supervised grasping
Robot learning in the wild?
Robot learning in homes
Large scale robot learning
Rich history
Under the hood
Deformable Object Manipulation
Approach
Key Challenges
Solutions
Conditional Policy Learning with MVP
Real Robot Experiments
Why forward models?
Learning visual forward models
Learning predictive representations
Contrastive representations
Contrastive forward models
One step MPC
Quantitative evaluation