Robotic systems are not only becoming increasingly complex, but so are the environments for which they are being built. While this poses various challenges for hardware design, the greater difficulties lie in devising appropriate control structures to move robots in these environments and creating artificial intelligence systems capable of solving the information processing problems posed by the robots' tasks. With new fields such as Autonomous Driving and Human Machine Interaction emerging in recent years, this development is accelerating, both in the academic and industrial sector.
The last decades of AI research have seen various predictions of when and by which means general AI systems would eventually emerge, and in the path towards them, powerful narrow AI systems capable of solving the kinds of tasks envisioned for future autonomous mobile agents. Yet despite the growing flexibility and an abundance of machine learning methods, with more algorithms, libraries and use cases published every year, software driving current robots is still mainly being created and tuned manually by experienced engineers using prior knowledge, as AI systems of said capabilities are still not available. Even the field of Deep Learning (DL), leading to major advances in areas such as machine vision, and thus to a revival of the use of artificial neural networks (ANNs), is slowing down in development again, with publications showing smaller improvements on more specific problems and increasingly revealing the limitations of this technology.
It is easy to attribute the difficulty to produce more advanced AI systems to the inadequacy of available computing power. However a look at nature might indicate otherwise, given that even insects possessing a mere few hundred thousand neurons show abilities still beyond any current machine learning system. This fact, together with observations such as the decreasing gains in performance of DL systems when being trained with increasingly large sets of data, or evolutionary algorithms (EAs) usually reaching local optima of comparably simple solutions, hints at it still being our methods rather than our resources limiting the development of more powerful AIs.
This talk will explore the differences between the biological processes that led to the evolution of a broad spectrum of intelligence in nature and current methods in robotic and AI design, machine learning and evolutionary algorithms. Among the questions being discussed are (i) 'Is it impossible to separate the development of artificial intelligence and control structures from the hardware structures of robots and still develop intelligence as capable as seen in animals?' and (ii) 'Are our current abstractions of biology such as ANNs and EAs adequately representing the mechanisms governing neurological and evolutionary development in nature?'