Angela Schoellig

University of Toronto
Project Title: Safe learning-based control for high-precision assembly robots in advanced aerospace manufacturing
Industry Partner: MDA Corporation
Platforms: Agile Computing, Large Memory System, Cloud Analytics

Advanced Manufacturing Health

Ontario’s aerospace industry has a distinguished history of innovation, and it provides a strong contribution to the Canadian economy with its more than 350 firms employing more than 20,000 Ontarians. Our industrial partner, MDA Robotics and Automation, is a leading company in providing advanced engineering solutions for various industries, including aerospace robotics and manufacturing. Having designed and built the famous Canadarm and Canadarm2 for manipulation in space, building versatile, high-precision assembly robots for aerospace applications on earth (such as airframe assembly for an airplane) is at the top of the company’s current priorities. Potential advantages include lowering the assembly cost, increasing the productivity, and reducing the risk of human error. Unlike high-volume manufacturing processes, for which it pays off to go through the lengthy, possibly month-long, process of manually programming and tuning the robot controls to carry out a limited range of repetitive tasks in a well-defined environment, the diversity of tasks and the uncertainty of the work environment in aerospace manufacturing motivate the need for smarter robots that can safely interact with their unpredictable environment, improve their performance through learning, and more importantly, generalize the knowledge from previous tasks and/or other robots to adapt to new tasks not trained on before. The objective of the project is to combine advanced methods from control theory, machine learning, and optimization to develop computationally efficient, learning-based control algorithms that improve the assembly robots’ performance in uncertain scenarios. Our strategy includes utilizing (i) well-developed algorithms in our group for online, safe learning of robots, and (ii) deep learning for transferring the knowledge learned by one robot in a particular task to other similar tasks and/or robots.