Applying Deep Reinforcement Learning to Cable Driven Parallel Robots for Balancing Unstable Loads: A Ball Case Study

Abstract

The current pandemic has highlighted the need for rapid construction of structures to treat patients and ensure manufacturing of health care products such as vaccines. In order to achieve this, rapid transportation of construction materials from staging area to deposition is needed. In the future, this could be achieved through automated construction sites that make use of robots. Toward this, in this paper a cable driven parallel manipulator (CDPM) is designed and built to balance a highly unstable load, a ball plate system. The system consists of eight cables attached to the end effector plate that can be extended or retracted to actuate movement of the plate. The hardware for the system was designed and built utilizing modern manufacturing processes. A camera system was designed using image recognition to identify the ball pose on the plate. The hardware was used to inform the development of a control system consisting of a reinforcement-learning trained neural network controller that outputs the desired platform response. A nested PID controller for each motor attached to each cable was used to realize the desired response. For the neural network controller, three different model structures were compared to assess the impact of varying model complexity. It was seen that less complex structures resulted in a slower response that was less flexible and more complex structures output a high frequency oscillation of the actuation signal resulting in an unresponsive system. It was concluded that the system showed promise for future development with the potential to improve on the state of the art.

Continuous Jumping of a Parallel Wire-Driven Monopedal Robot RAMIEL Using Reinforcement Learning

“…Reinforcement learning has also been introduced to some wire-driven robots [12]- [14]. On the other hand, [12], [14] is applied only to simulations, and [13] deals only with a simple and relatively static motion of walking with a fixed upper end. In this study, we use reinforcement learning to realize dynamic and difficult jumping motions in a wire-driven robot in simulation and on the actual robot, which are more difficult than those in previous studies on manipulators, quadruped robots, etc.…”

“…Some of them have been applied not only to simulations but also to actual robots [10], [11]. Reinforcement learning has also been introduced to some wire-driven robots [12]- [14]. On the other hand, [12], [14] is applied only to simulations, and [13] deals only with a simple and relatively static motion of walking with a fixed upper end.…”