About Decentralized Reasoning in Reduced Information Spaces
As robotics evolves, so do the challenges that the researchers face. A decade ago the research challenge was to enable navigation from one room to another in static, fully-known environments, while today vehicles are capable of autonomously navigating large-scale semi-structured and complex outdoor environments (DARPA Urban Challenge, DARPA UPI/Crusher Programs). Researchers have achieved major advances in both hardware and algorithmics, leading to new research challenges to enable robots to operate fully autonomously in the real-world, on a par with people, and as part of large static and mobile networks. Firstly, robots now have access to an array of high quality sensors such as 3D LADAR scanners, high resolution cameras, and high quality inertial navigation. As a result, robots know their own location, as well as those of inanimate objects and people even at moderate distances. However, this glut of sensor data implies that fast, intelligent agent action requires reducing information to what is relevant and actionable. Secondly, mobile robots now both move and manipulate objects in the human environment. As a result, the operational space of the robot became much larger and much less controlled, further increasing the size of the information spaces. Planning algorithms need to be able to project these operational spaces onto much smaller information spaces carrying salient features; this projection function may need to occur as a function of a task and the prior history. Thirdly, as robots become more au-
tonomous, they also need to interact with other agents, especially people, and be able to recognize and understand intent and plans. Doing so efficiently requires compressing large quantities of data about observed behavior over time into an understanding of the behaviors of other agents.
The core challenge is the need to reason quickly and effectively in huge information spaces.
The key problems that must be addressed in terms of autonomy—from perception, to recognition, to planning under uncertainty, to adapting to changing environments, to coordinating different agents—can be understood as acquiring, compressing, and planning with information states. In this project, Decentralized Reasoning in Reduced Information Spaces, we are developing a new science of reduced information spaces.
Our effort will leverage new computational ideas to “project” full information states onto much smaller, reduced information spaces and to reason in these reduced spaces. Our goals are to develop techniques and architectures for:
a) automatic reduction and low-dimensional approximation of information spaces
b) reasoning in reduced and approximate information spaces
c) decentralized decision- making and coordination in reduced and approximate information spaces.