Learning by Demonstration for Planning Activities of Daily Living in Rehabilitation and Assistive Robotics

This letter presents a motion planning system for robotic devices to be adopted in assistive or rehabilitation scenarios. The proposed system is grounded on a learning by demonstration approach based on dynamic movement primitives (DMP) and presents a high level of generalization allowing the user to perform activities of daily living. The proposed approach has been experimentally validated on a robotic arm (i.e., the Kuka LWR4+) attached to a human subject wrist. Two experimental sessions have been carried out in order to: 1) evaluate the differences between our approach and the one proposed in “Dynamical movement primitives: Learning attractor models for motor behaviors” (A. J. Ijspeert et al., Neural Comput., 2013) in terms of reconstruction error between the demonstrated trajectory and the learned one, and in terms of memory size required to record the database of DMP parameters; and 2) measure the generalization level of the proposed system with respect to the variation of the object positions by evaluating the success rate of the task execution. The experimental results demonstrate that the proposed approach allows 1) reproducing the user's personal motion style with high accuracy and 2) efficiently generalizing with respect to the change of object position. Furthermore, a significant reduction of memory allocation for the database can be achieved, with a consequent significant computational time saving.