Robotic rock classification and autonomous exploration.
详细信息
文摘
The capability to autonomously conduct a scientific investigation is essential to the next generation of robotic vehicles for the exploration of the solar system. Bandwidth and storage limitations mean that current vehicles return a fraction of the scientific data they are capable of acquiring, and cannot respond to observations in real time. Future missions, such as a robotic search for life under the ice cap of Europa will have extremely limited communications with Earth, yet operate in unknown and dynamic environments. Such a robot must autonomously recognize scientifically interesting occurrences and conduct follow up observations as required.;A first step in the development of an autonomous science capability for a robotic explorer is the capability to autonomously recognize a restricted class of scientifically interesting objects (such as rocks, meteorites or fossils) and survey an area, learning what objects are present and permitting the identification of anomalies.;This thesis is about the application of Bayesian statistics to the problem of autonomous geological exploration with a robotic vehicle. It concentrates on the sub-problem of classifying rock types while addressing the issues associated with operating onboard a mobile robot, and argues that the Bayesian statistical paradigm, using a Bayesian network generative statistical model of the geological environment, is particularly suited to this task. This paradigm is extended in a natural way to solve the more general robotic problems of autonomously profiling an area and allocating scarce sensor resources.;Major considerations are the need for multiple sensors and the acquisition of sensor data at different geographic locations. Needless sensor use should be curtailed, such as when an object is sufficiently well identified by sensor data acquired so far. By investigating rocks in many locations, the robot has the opportunity to profile the environment. Different rock samples are statistically dependent on each other. These dependencies can be exploited to substantially improve classification accuracy.;The classification system has been implemented onboard the Nomad robot developed at Carnegie Mellon University, and applied to the task of recognizing meteorites amongst terrestrial rocks in Antarctica. In January 2000 A.D., Nomad was deployed to Antarctica where it made the first autonomous robotic identification of a meteorite.