Distributed Robust Inference and Data Association

We consider multi-robot inference over variables of interest (e.g. robot poses), from unknown initial robot poses and undetermined data association. This problem is relevant for different multi-robot collaborative applications, such as cooperative mapping, localization, tracking, and surveillance. Collaboration requires the robots to share a common world model and to be able to correctly interpret information communicated with each other. We show that establishing this collaboration first requires inferring concurrently a common reference frame between the robots and resolving data association. The problem becomes even more challenging in the incremental setting and in presence of perceptual aliasing (e.g. two different but similar in appearance corridors). We develop an Expectation-Maximization (EM) and model selection framework to address this problem.

Below is a demonstration of our approach using real-world experiment involving two quadrotors (colored blue and red) operating in indoor environment and sharing informative laser scans. As the robots do not have a common reference frame established, their initial poses are set to /arbitrary/ values; in practice, both robots start operating from the same location (middle image). Multi-robot candidate correspondences are generated by ICP-matching the shared laser scans. After some time, our approach successfully estimates the initial relative pose between the robots and determines multi-robot data association (right image). From that moment, it becomes possible for the robots to robustly infer variables of interest, in this case each other’s trajectories, and to identify the inlier correspondences (denoted in black) in newly arriving data. The movie shows the process from the perspective of each robot (first red and then blue robot).

A distributed real-time implementation of the approach is demonstrated below.

Related Publications: link to bib file