SegVisRL: Visuomotor Development for a Lunar Rover for Hazard Avoidance using Camera Images
Authors:
Tamir Blum,
Gabin Paillet,
Watcharawut Masawat,
Mickael Laine,
Kazuya Yoshida
Abstract:
The visuomotor system of any animal is critical for its survival, and the development of a complex one within humans is large factor in our success as a species on Earth. This system is an essential part of our ability to adapt to our environment. We use this system continuously throughout the day, when picking something up, or walking around while avoiding bumping into objects. Equipping robots w…
▽ More
The visuomotor system of any animal is critical for its survival, and the development of a complex one within humans is large factor in our success as a species on Earth. This system is an essential part of our ability to adapt to our environment. We use this system continuously throughout the day, when picking something up, or walking around while avoiding bumping into objects. Equipping robots with such capabilities will help produce more intelligent locomotion with the ability to more easily understand their surroundings and to move safely. In particular, such capabilities are desirable for traversing the lunar surface, as it is full of hazardous obstacles, such as rocks. These obstacles need to be identified and avoided in real time. This paper seeks to demonstrate the development of a visuomotor system within a robot for navigation and obstacle avoidance, with complex rock shaped objects representing hazards. Our approach uses deep reinforcement learning with only image data. In this paper, we compare the results from several neural network architectures and a preprocessing methodology which includes producing a segmented image and downsampling.
△ Less
Submitted 26 March, 2021;
originally announced March 2021.
