How the robots could develop their motor intelligence? We have seen a remarkable evolution in robotics such as the Atlas robot performing gymnastics routines. However, despite the progress, these behaviors are typically scripted or generated using domain-specific knowledge. These methods highly rely on static models and human’s intuition that limit its application to a broader domain. Furthermore, recent advance in deep learning has shown impressive results, but unfortunately, it builds a narrow intelligence that cannot be transferred to different circumstances. This kind of intelligence is not reliable and cannot be robustly generalized in motor control. Indeed, it is naively approached as a ‘‘universal’’ function approximator, and the lack of building abstract concepts affects its generalization.
In our group, we focus on creating motor intelligence in legged robots. Our aim is to find computational principles that describe what we call the motor function. Such function should have a conceptual structure, and not only a universal approximator. With this, our robots could
To achieve so, we study a range of different disciplines: numerical optimization, model predictive control, deep learning, robot design, and perception. Our research agenda is devoted to enable agile locomotion and manipulation in robots with legs and arms.
How to achieve Athletic intelligence?
— Carlos Mastalli (@CarlosMastalli) December 9, 2020
In a nutshell, I believe that combing explainable control approaches, efficient numerical optimization aspects, and benefits of data-driven approaches is crucial!@EdinRobotML @turinginst @MemmoEU pic.twitter.com/JmQkfrv5Ku