Learning by Demonstration with Baxter Humanoid

Abstract

Despite robots’ high capabilities to perform various tasks, they require programming for each new task. Learning by demonstration (like humans) enables robots to perform new tasks without the need for well-crafted programs. The aim of this project is to apply learning by demonstration method in which the user performs two simple tasks in front of the robot and the robot imitates them. Baxter humanoid is used in the project; it has two arms with seven joints on each arm, a face display and a non-programmable wheeled base. The task inspired from studies on animal cognition to use a tool appropriately in order to fetch/obtain a reward, following the demonstration of the human. The overall learning by demonstration system integrated two core sensory-motor loops: (1) Action Observation: to observe/represent user demonstration, the VICON system in the robotics arena was integrated to the Baxter so as to receive information about objects in the scene, movement trajectory, ensuing consequences; (2) Action Generation: to make the robot reproduce the observed demonstration of the tool use. Baxter imitated the user’s demonstrations successfully and the required results achieved. The developed system is in general task agnostic, and can be further extended in the domain of motor primitives by moving from trajectories to movement shapes. Further, the system is useful in industrial assembly tasks since the robot does not need programming for each new task which saves time and simplifies robot work.