Controller design for high-dimensional bimanual multi-fingered hands

Develop reliable control design methodologies for multi-fingered bimanual robot hands operating in high-dimensional action spaces that address challenges arising from mechanical design complexity and actuation reliability.

Background

The paper reviews prior work on bimanual manipulation and emphasizes the difficulty of coordinating high-dimensional action spaces inherent to multi-fingered hands. While multi-fingered hands are theoretically appealing for bimanual systems, the practical design of controllers is hindered by mechanical and actuation constraints.

This open problem motivates the authors’ approach of using deep reinforcement learning and sim-to-real transfer to learn control policies rather than hand-crafted controllers.