Learning the end-effector pose from demonstration for Bionic Handling Assistant robot

For most of the rigid manipulators, it is possible to apply a gravity compensation mode, by which the user is able to easily reconfigure the arm and record the necessary data. However, due to the specific characteristics of soft robots such as elastic properties and complex dynamics, it is usually very difficult to implement kinesthetic teaching for Learning from Demonstration (LfD) scenarios. This paper tackles this problem on a soft continuum robot named Bionic Handling Assistant (BHA). We propose to use an active compliant controller that facilitates the kinesthetic teaching for the user while recording the position and orientation of the robot’s end-effector. The recorded demonstrations are then encoded with a task-parameterized probabilistic model through two separate dynamical systems (one for the position and one for the orientation). The approach was evaluated by conducting two experiments on the BHA robot.