In modern society, robots are increasingly used to handle dangerous, repetitive and/or heavy tasks with high precision. Because of the nature of the tasks, either being dangerous, high precision or simply repetitive, robots are usually constructed with high torque motors and sturdy materials, that makes them dangerous for humans to handle. In a car-manufacturing company, for example, a large cage is placed around the robot’s workspace that prevents humans from entering its vicinity. In the last few decades, efforts have been made to improve human-robot interaction. Often the movement of robots is characterized as not being smooth and clearly dividable into sub-movements. This makes their movement rather unpredictable for humans. So, there exists an opportunity to improve the motion generation of robots to enhance human-robot interaction. One interesting research direction is that of imitation learning. Here, human motions are recorded and demonstrated to the robot. Although the robot is able to reproduce such movements, it cannot be generalized to other situations. Therefore, a dynamical system approach is proposed where the recorded motions are embedded into the dynamics of the system. Shaping these nonlinear dynamics, according to recorded motions, allows for dynamical system to generalize beyond demonstration. As a result, the robot can generate motions of other situations not included in the recorded human demonstrations. In this dissertation, a Reservoir Computing approach is used to create a dynamical system in which such demonstrations are embedded. Reservoir Computing systems are Recurrent Neural Network-based approaches that are efficiently trained by considering only the training of the readout connections and retaining all other connections of such a network unchanged given their initial randomly chosen values. Although they have been used to embed periodic motions before, they were extended to embed discrete motions, or both. This work describes how such a motion pattern-generating system is built, ...
Adaptive control of compliant robots with Reservoir Computing
2013-01-01
Hochschulschrift
Elektronische Ressource
Englisch
DDC: | 629 |
NTRS | 1990
|Compliant Floating-Base Control of Space Robots
Deutsches Zentrum für Luft- und Raumfahrt (DLR) | 2021
|BASE | 2019
|