Rethinking Dexterous Manipulation: From Algorithms to Hardware

During the last decade, learning-based techniques have been quite successful in generating motor skills in simulations. However, these techniques in their current form are less effective on real robots, especially in contact-rich dexterous manipulation settings. In this talk, I'll revisit the problem of learning dexterous manipulation from the first principles. I'd argue that in addition to algorithmic developments, learning paradigms in the real world, as well as hardware infrastructure, needs significant attention to imparting human-level dexterity to our robots in a scalable way. 

On the algorithmic front, I'll discuss a game-theoretic formulation for model-based reinforcement learning (MBRL) that not only unifies, and generalizes many previous MBRL algorithms but also provides guidelines for designing more stable algorithms capable of learning general manipulation behaviors that can be retargeted to new unseen tasks. I’ll also share (less discussed) experiences and lessons learned while maturing our algorithmic paradigms to acquire high-dimensional, contact-rich, dexterous manipulation behaviors in the real world in a scalable way. And how we went from half a million-dollar hardware that broke every 30 mins to a scalable and modular 5000$ setup that can learn behaviors unattended for weeks without any human intervention.