Towards learning robotic manipulation and tool use

Ville Kirki
(Aalto University, Finland)


Manipulation capabilities have recently become one central focus of research in robotics. In particular, the question how to provide capabilities beyond grasping in standard pick-and-place operations has gained increasing attention, for example, how to provide robust skills for using door handles or tools. While the main problem in robotic grasping is how to achieve and maintain a rigid relationship between a robot hand and a target object, tool use requires control of interaction forces between objects. In this talk I will look at some of our recent results how learning can be used to provide such skills.

Learning manipulation skills can be viewed from three perspectives: skill acquisition, skill adaptation and skill generalization. In skill acquisition the challenge is to build a robust primitive skill. Skill acquisition is often done by transferring a skill from a human to the system. The process, called programming by demonstration or imitation learning, is usually solved as a supervised learning problem. The transferred skills can further be adapted to new contexts, for example by reinforcement learning. Finally, the ultimate goal of learning is to provide generalizable skills that can be applied in different contexts. I will review our recent work that aims to build these generalizable skills through a combination supervised and reinforcement learning.