Simulating DuCTT and optimizing control for DuCTT with machine learning

摘要

Air duct inspection frequently requires mobile robots to visit areas inaccessible to humans. Tensegrity robots, with their small mass and cross-section are highly suited to inspecting air ducts without impeding the flow within the duct. One tensegrity robot designed at the UCSD Coordinated Robotics Lab for this task is the Duct Climbing Tetrahedral Tensegrity (DuCTT) robot. This robot consists of two tetrahedral sections connected by eight cables. This work presents a way to simulate this robot in the NASA Tensegrity Robotics Toolkit (NTRT). Once the robot is simulated, control strategies can be explored in a variety of different environments. These strategies can get the robot to climb and traverse air ducts. and ways to optimize controllers based on those strategies. The two different strategies are using sine waves to control the actuators, and a state-machine controller. Each controller is optimized using one stage of Monte-Carlo parameter estimation and a second stage genetic algorithm to improve the parameters found by the first stage. The state-machine controller ends up with better performance, 6.77 cm/s, compared to the sine wave controller's 0.38 cm/s