EVENTS

Despite significant advances in AI, robots still struggle with tasks involving physical interaction. Robots can easily beat humans at board games such as Chess or Go but struggle to skillfully move the game pieces by themselves (the part of the task that humans subconsciously succeed in). Learning manipulation skills is both hard and fascinating because the movements and behaviors to acquire are tightly connected to our physical world and to embodied forms of intelligence.

I will present an overview of representations and learning approaches to help robots acquire manipulation skills by imitation and self-refinement. I will present the advantages of targeting a frugal learning approach, where the term “frugality” has two goals: 1) learning manipulation skills from only few demonstrations or exploration trials; and 2) learning only the components of the skill that really need to be learned.

Toward this goal, I will emphasize the roles of geometry, manifolds, implicit shape representations and distance fields as inductive biases to facilitate human-guided manipulation skill acquisition. I will also show how ergodic control can provide a mathematical framework to generate exploration and coverage movement behaviors, which can be exploited by robots as a way to cope with uncertainty in sensing, proprioception and motor control.

Institutions

• AI for Good