Study on Reward Multidimensionality in Deep Reinforcement Learning of Spherical Synchronous Actuators
Authors:
- Kazuki ShorikiKSKazuki Shoriki
- Akira HeyaAHAkira Heya
- Katsuhiro HirataKHKatsuhiro Hirata
Open Access
- 1 January 2022
- journal article
- research article
- Published by Japan Society of Applied Electromagnetics and Mechanics in Journal of the Japan Society of Applied Electromagnetics and Mechanics
- Vol. 30 (2) , 185-190
Abstract
The spherical synchronous actuator can be driven in multiple-degrees-of-freedom by a single unit. This simplifies the mechanism and is expected to be applied to industrial robots. The position of the actuator is controlled using a torque map which is a torque data at each posture. However, there is a problem that the positional accuracy is degraded by measurement errors in the torque map and friction. To solve this problem, we propose a novel attitude control system for the spherical actuator using a deep reinforcement learning algorithm. The simulation results showed that the control system was successfully implemented without a torque map. The system was able to follow the step response within 0.1 seconds, and the steady-state deviation was less than 0.1 degrees.Keywords
This publication has 3 references indexed in Scilit:
- Improvement of Analysis Accuracy in Torque Calculation of Multi-Degree-of-Freedom Spherical ActuatorsJournal of the Japan Society of Applied Electromagnetics and Mechanics, 2021
- A Nonlinear Time Scale Translation for Input Torque Patterns of a Robot Manipulator Considering the DynamicsTransactions of the Society of Instrument and Control Engineers, 1996
- Bettering operation of Robots by learningJournal of Robotic Systems, 1984