Abstract

Remotely controlled humanoid robots have countless applications in both industry and science. They can be effectively controlled by making them recreate the pose of a motion-tracked human operator in real-time. The robots often have industrial closed-loop controllers in their joints to perform their movements. PID controllers are among the most common controllers, but they are notoriously hard to tune. This thesis aims to provide tools to assist with PID tuning for both experienced and inexperienced users. It focuses on the specific use-case of remote operation of humanoid robots. To achieve this, a tuning framework consisting of tools, APIs, and UIs is developed. It integrates into an existing remote operation solution, which allows operators to control a virtual, simulated humanoid robot. The tuning framework facilitates both manual and automatic PID tuning. It provides functionality for data collection, motion playback, and performance evaluation. The automatic tuning feature offers multiple heuristics, which are tested on four different robot setups. Their performance is compared to a manual reference tuning. The top heuristics for each setup are identified based on multiple performance metrics. The thesis concludes that the manual tuning generally outperforms all heuristics of the automatic tuning routine. The development of new, purpose-built heuristics for the specific use-case is suggested. Alternative methods of robot control and PID tuning are proposed in the outlook.

Results

For detailed results, please see the attached thesis.