Adaptive rover behavior based on online empirical evaluation: Rover–terrain interaction and near-to-far learning
Article first published online: 23 DEC 2009
Copyright © 2009 Wiley Periodicals, Inc.
Journal of Field Robotics
Special Issue: Vehicle–Terrain Interaction for Mobile Robots
Volume 27, Issue 2, pages 158–180, March/April 2010
How to Cite
Krebs, A., Pradalier, C. and Siegwart, R. (2010), Adaptive rover behavior based on online empirical evaluation: Rover–terrain interaction and near-to-far learning. J. Field Robotics, 27: 158–180. doi: 10.1002/rob.20332
- Issue published online: 1 FEB 2010
- Article first published online: 23 DEC 2009
- Manuscript Accepted: 2 NOV 2009
- Manuscript Received: 4 JAN 2009
Owing to the fundamental nature of all-terrain exploration, autonomous rovers are confronted with unknown environments. This is especially apparent regarding soil interactions, as the nature of the soil is typically unknown. This work aims at establishing a framework from which the rover can learn from its interaction with the terrains encountered and shows the importance of such a method. We introduce a set of rover–terrain interaction (RTI) and remote data metrics that are expressed in different subspaces. In practice, the information characterizing the terrains, obtained from remote sensors (e.g., a camera) and local sensors (e.g., an inertial measurement unit) is used to characterize the respective remote data and RTI model. In each subspace, which can be described as a feature space encompassing either a remote data measurement or an RTI, similar features are grouped to form classes, and the probability distribution function over the features is learned for each one of those classes. Subsequently, data acquired on the same terrain are used to associate the corresponding models in each subspace and to build an inference model. Based on the remote sensor data measured, the RTI model is predicted using the inference model. This process corresponds to a near-to-far approach and provides the most probable RTI metrics of the terrain lying ahead of the rover. The predicted RTI metrics are then used to plan an optimal path with respect to the RTI model and therefore influence the rover trajectory. The CRAB rover is used in this work for the implementation and testing of the approach, which we call rover–terrain interactions learned from experiments (RTILE). This article presents RTILE, describes its implementation, and concludes with results from field tests that validate the approach. © 2009 Wiley Periodicals, Inc.