Improving perception of unknown environments or objects for robotic systems

点击翻译按钮获取中文摘要

Robotic systems were deployed, at first, in tightly controlled environments, such as laboratories or factory floors. Their operating strategies were based on pre-programmed knowledge, under the assumption that the environment was fully observable and would not change over time. Nowadays, robotic systems are being deployed increasingly in uncontrolled environments, such as buildings occupied by humans or outdoors. Robotic systems still need to perceive their environment in a useful manner, despite the fact that they might be operating in an unfamiliar environment. Consequently, these systems must extract and assimilate new knowledge on-the-fly, based on the information collected through their sensors, such as tactile sensors, cameras or laser range finders. Reliance on pre-programmed knowledge is no longer sufficient for these robots. The long term research objective of our program is to develop autonomous learning strategies for robotic systems interacting with the physical world, with an emphasis on the training of artificial perception systems. This applies to mobile robots operating in outdoor environments, or to industrial robots specialized in object recognition. Our research program is focused on two complementary issues: improving sensing capabilities and developing systems that learn on their own. We are looking, for example, at improving and exploiting tactile information for object recognition and terrain identification in robotics. Tactile information is very rich, as is confirmed with our everyday experience when manipulating objects with our hands. And by developing strategies on how to learn on their own, these systems will be able to quickly adapt to unforeseen conditions, a true hallmark of intelligent behavior. It is expected that our research will lead to a number of industrial applications. For example, we foresee applications of tactile sensing in agricultural robotics. Indeed, it is easier to discern a fruit from the surrounding stems and leaves, if one can incorporate both tactile and visual information.