JOMAR: Joint Operations with Mobile Autonomous Robots

Abstract

Abstract : Under this grant, we formulated and implemented a variety of novel algorithms that address core problems in multi-robot systems. These contributions roughly fall into two categories: state estimation and mapping, and robot perception and computer vision. 1. State Estimation and Mapping: * Methods for performing non-linear optimization with non-Gaussian error models. This provides a fundamental advantage over Gaussian methods which are unable to model real world sensor failure modes. This MaxMixture formulation is the standout success of this grant based on adoption and citation by the community. * A characterization of Global Positioning System (GPS) noise models in the MaxMixture framework, allowing significant improvements in GPS-aided navigation. * A data-association algorithm with applications to target tracking and computer vision applications, named the Incremental Posterior Joint Compatibility (IPJC) test, which computes optimal data associations in a small fraction of the time required by previous methods. 2. Robot Perception and Computer Vision: * A method for learning visual features based on the needs of an application. Previous approaches rely on humans to design high-performance visual features; we show for the first time that such filters can be learned in-situ. * A new camera calibration system that achieves dramatically more accurate and consistent calibration results than previous methods. 3. Radio Communication and Mesh Networking: New methods for predicting the signal strength between two robots in a mesh network leveraging both previous robot radio communication attempts and non-radio sensor data such as LIDAR.