Distributed and Modular Bio-Inspired Architecture for Adaptive Motor Learning and Control

Recent studies have demonstrated that autonomous robots can outperformthe task they are programmed for, but are limited in their ability to adapt to unexpected situations (Ingrand and Ghallab, 2017). This limitation is due to the lack of generalization, i.e., the robot can not transfer knowledge across multiple situations.

Even the application of modern artificial intelligence (AI) techniques does not support a robust generalization when the range of probable inputs is infinite (Yang et al., 2018; Mnih et al., 2015; Cai et al., 2017;Kober et al., 2013). As a matter of fact, AI methods can interpolate knowledge but not extrapolate it, i.e., they can adapt on new, unseen data that are within the bounds of their experience, but not on data that are outside the bounds.So far, robots have been mostly treated as stand-alone systems in a vacuum,while the real world is more complex and includes continuous interaction with external entities.

Accordingly, the design of a generalized robotic controller is not trivial, in particular when the dynamical condition are unknown.