Helping robots learn new skills quickly and independently
An EU-funded project is looking to increase the capacity of robots to set themselves goals and develop skills autonomously. This could boost EU competitiveness in the field of artificial intelligence and advance understanding of learning processes.
© Optinik, #200630153, 2019. source: stock.adobe.com
Aimed at bringing about a paradigm shift in the development of autonomous learning robots, the EU-funded GOAL-ROBOTS project is based on two insights. First, to undertake autonomous, open-ended learning, robots must be able to identify goals and hence to generate the tasks they need to perform to accomplish them. Secondly, new learning algorithms can use self-generated goals to accelerate skills development.
GOAL-ROBOTS builds on this by developing computational architectures and algorithms that support two functions: self-generation of goals without the assistance of humans, based on things like curiosity, and through use of the goals to autonomously learn large repertoires of skills. In this context, when new skills have been acquired, any similarity between goals can speed up further skills learning.
Furthermore, the project aims to increase knowledge of how goals are formed in humans and how they underpin childrens learning. Experiments involving 50 infants aged 3-18 months focus on how the processing of events arising from their actions drives childrens acquisition of skills. The findings are incorporated into the projects computational architecture for transfer to robots.
Four increasingly complex demonstrations with robots show how the algorithms and architectures allow the self-generation of goals and the autonomous development of skills and their later use for solving tasks defined by the potential users of the robots.
The first demonstration shows the basic open-ended learning process; the second shows the acquisition of multiple skills on the basis of self-generated goals; while the third shows the acquisition of skills in varying conditions. The fourth demonstration builds on the previous three to allow the robot to solve tasks defined by the user and involving different objects.
By enabling robots to acquire flexible skills autonomously and in conditions that are unforeseen at the time of their design, GOAL-ROBOTS represents a breakthrough in the field of autonomous robotics. This allows the robots subsequently to be able to deploy the goals and skills acquired autonomously to perform user-defined tasks with little or no additional learning.