Learning Robotic Tasks from Human Demonstrations

Many tasks in manufacturing, especially in assembly, are highly skill intensive and require many years of training to perfect in some instances. At the same time, some of these tasks can be dirty, dangerous and demeaning. While automation has had key successes in overcoming this challenge, this has largely been the case for repetitive, relatively simple tasks. Highly skilled tasks often rely on a combination of integrated sensing, cognition and dexterity. Especially the cognitive decision-making processes that connect sensing with dexterous manipulation are very hard to programme using traditional robot programming methods. Yet, it is essential for industrial systems to have an increased level of cognitive autonomy to better cope with variations and less repetitive processes for instance during the final assembly of cars, aircraft and other assembled products.
This project will investigate how emerging developments in teaching robots from human demonstrations of tasks can be used to reduce or even remove the need for programming. It is expected that this project will work with methods to capture human motion and other process parameters such as interaction forces but also eye tracking. The challenge will be to learn the underlying skills operators use to complete different tasks and transfer them to a robotic system. The ambition is that a knowledge base of skills can be build up that can be transferred and improved between different industrial robots. Virtual Reality could play a key role in training robots in the future for new tasks before deploying them into a factory ready to complete tasks with human like adaptation capabilities.