Anelia Angelova

Slip Prediction from Visual Information for Autonomous Robots Anelia Angelova (Caltech), Larry Matthies (JPL), Daniel Helmick (JPL), Pietro Perona (Caltech)

We consider the problem of slip prediction from a distance for ground wheeled robots, using stereo imagery as input and learning from previous examples of traversing similar terrains. A generic nonlinear regression framework is proposed in which the terrain type is determined from appearance and then a nonlinear model of slip is learned for a particular terrain type.

To achieve fully autonomous behavior, the vehicle's mechanical sensors are used as supervision to learning a vision-based terrain classification. We present a probabilistic framework in which the visual information and the mechanical supervision interact to learn particular terrain types and their properties. This framework is further extended to handle high dimensional input spaces. Our experiments on several off-road terrains show that using mechanical measurements as automatic supervision improves the visual-based classification and approaches the results of learning with manual supervision. This research is motivated by the need to assess the risk of the rover getting trapped before entering a particular terrain. The proposed algorithms will enable the rover to drive safely on slopes, learning autonomously about different terrains and their slip characteristics.

We further propose a learning framework which enables fast recognition of terrains. Instead of building a monolithic classifier with uniformly complex representation for each class, the main idea here is to actively consider the labels or misclassification cost while constructing the classifier. The algorithm automatically builds a variable-length visual representation which varies according to the complexity of the classification task. This enables faster recognition of terrain types during testing.