International Journal of Robotics, Theory and Applications

International Journal of Robotics, Theory and Applications

Fast Reinforcement Learning approach to robust optimal control of bipedal robots with Point-feet

Document Type : Original Article

Authors
Majid Anjidani 1
Mohammad Reza Jahed Motlagh 2
Mahmood Fathy 3
Majid Nili Ahmadabad 4
1 PhD student, Iran University of Science and Technology, Tehran, Iran
2 Faculty member of computer department, Iran University of Science and Technology, Tehran, Iran
3 Faculty member of computer dep., Iran University of Science and Technology, Tehran, Iran
4 Faculty member of Electrical and Computer Department, University of Tehran, Tehran, Iran
Abstract
Designing a walking gait for biped robots, that can preserve stability against a known range of disturbances, is very important in real applications. In the area of biped robots with point-feet, designing an exponentially stable walking gait with desired features has been recently done by an online reinforcement learning method called PI^2-WG. However, the designed gait might not be robust enough against disturbances. Therefore, we extend a robust version of PI^2-WG to design an exponentially stable walking gait which is robust against modeling errors/disturbances and we call it R-SPI^2-WG. It is done by minimizing the costs of worst rollouts which are generated in presence of different modeling errors/disturbances. We study the ability of the proposed method to adapt the controller of RABBIT, which is a planar biped robot with point-feet, for some robust applications. The simulation results show that the designed gaits are exponentially stable and robust against modeling errors/disturbances in a feasible range.
Keywords
Legged robots
reinforcement learning
robust gait optimization

  • Receive Date 11 December 2016
  • Revise Date 27 January 2017
  • Accept Date 18 February 2017
  • First Publish Date 01 June 2021