Adaptation and coaching of periodic motion primitives through physical and visual interaction

Abstract

In this paper we propose and evaluate a control system to 1) learn and 2) adapt robot motion for continuous non-rigid contact with the environment. We present the approach in the context of wiping surfaces with robots. Our approach is based on learning by demonstration. First an initial periodic motion, covering the essence of the wiping task, is transferred from a human to a robot. The system extracts and learns one period of motion. Once the user/demonstrator is content with the motion, the robot seeks and establishes contact with a given surface, maintaining a predefined force of contact through force feedback. The shape of the surface is encoded for the complete period of motion, but the robot can adapt to a different surface, perturbations or obstacles. The novelty stems from the fact that the feedforward component is learned and encoded in a dynamic movement primitive.By using the feedforward component, the feedback component is greatly reduced if not completely canceled. Finally, if the user is not satisfied with the periodic pattern, he/she can change parts of motion through predefined gestures or through physical contact in a manner of a tutor or a coach.The complete system thus allows not only a transfer of motion, but a transfer of motion with matching correspondences, i. e. wiping motion is constrained to maintain physical contact with the surface to be wiped. The interface for both learning and adaptation is simple and intuitive and allows for fast and reliable knowledge transfer to the robot.Simulated and real world results in the application domain of wiping a surface are presented on three different robotic platforms. Results of the three robotic platforms, namely a 7 degree-of-freedom Kuka LWR-4 robot, the ARMAR-IIIa humanoid platform and the Sarcos CB-i humanoid robot, depict different methods of adaptation to the environment and coaching.

Learning the Correct Robot Trajectory in Real-Time from Physical Human Interactions

“…For example, if the task requires significant interaction with the environment-such as sawing-collaborative approaches are more appropriate [14,35,38]. Local modification approaches become increasingly relevant if the task involves multiple repetitions: the robot can store and replay the local adjustments between iterations, enabling the end user to intuitively and iteratively correct the robot's trajectory [1,9,11,13,15,24,27]. For scenarios where the task has few features (i.e., two to three), global replanning methods have already been implemented online [4,5]: thus, in practice, the time spent replanning may not be significant enough to warrant our alternate approach.…”

“…Adapting existing trajectories. Some prior approaches update the robot's global trajectory by adapting its current trajectory to the new scenario [13,29,32,37]. In Nierhoff et al [32], for example, the authors adapt the robot's entire trajectory in real time to satisfy changing environment constraints while remaining close to the original trajectory (without physical interactions).…”

“…In Nierhoff et al [32], for example, the authors adapt the robot's entire trajectory in real time to satisfy changing environment constraints while remaining close to the original trajectory (without physical interactions). More related to our approach is work by Gams et al [13]: here, the robot learns from physical coaching during periodic movements. Unlike this research, our work focuses on trajectory adaptation with optimal control and does not require periodic movements or multiple trials.…”

“…In this section, we map the human's physical interactions u h into real-time updates of the estimate θ so that the robot can learn the right trajectory from physical corrections. This approach is inspired by our insight from Section 1 and prior work on learning from pHRI [1,4,13,19,27]: the human interactions reduce the error between the robot's current trajectory and their preferred trajectory. We define e = [e T p , e T v ] T , where e p ∈ R n is the position error and e v ∈ R n is the velocity error.…”

A Passivity-Based Strategy for Manual Corrections in Human-Robot Coaching

“…Nevertheless, refs. [30,31] implement robot coaching by using visual and force feedback to change the imitation-related error signal. If the user is not satisfied with the periodic pattern, he/she can change parts of the motion through predefined gestures or physical contact.…”