Incremental Learning for Robot Sensing and Control

This proposal addresses key open challenges identified during the LAGR program for the practical use of adaptive, vision-based robot navigation in commercial settings. First, the adaptive vision system learns quickly, but forgets as quickly. This will be addressed by using an ensemble of "expert" classifiers, each of which specializes for a particular environment and can be quickly activated when the environment matches its domain of validity. Second, a new type of cost map will be used which accumulates high-level feature vectors, rather than traversability values. A global cost map will also be integrated. Third, we will pre-train the convolutional net feature extractor using the latest unsupervised algorithms for learning hierarchies of invariant features. Fourth, the limited power of general-purpose CPUs will be lifted by using a highly compact, dedicated FPGA-based hardware platform to run computationally intensive parts of the system. Implementations on commercially available GPUs will also be explored. Finally, to achieve portability and modularity, we will make our implementation independent of a particular robot platform and support a wide range of sensor types including stereo cameras and LIDAR. The result will be a highly-compact, low-power, self-contained, low-cost, vision-based navigation system for autonomous mobile robots.