Teaching robots the use of human tools from demonstration with non-dexterous end-effectors

Abstract

Commercial, affordable and general-purpose robots like the PR-2, Baxter and UBR-1 robots can take over a wide range of tasks or assist human workers in a mixed human-robot environment. However, end-effectors on these robots are usually restricted to low-cost, non-dexterous grippers which constrains the application scenarios. We aim at increasing their range by teaching such robots the use of human tools by demonstration. We present a novel and compact model for the use of human tools and propose a dual-gripper strategy in replacement of the much less widely deployed dexterous hand for tool manipulation. Especially, we propose a hierarchical architecture to embed tool use in a learning from demonstration framework, learning temporal order for dual-arm coordination at higher level and Dynamic Movement Primitives at lower level for a multi-step execution. The approach is demonstrated and evaluated on a Baxter research robot for three human tools.