Robust Distributed 3D Mapping With Communication Constraints

Abstract

Rapid autonomous exploration of challenging, GPS-denied environments such as underground mines provides essential information to search and rescue as well as defense operations. We pursue a distributed perception strategy for a team of robots to develop a consistent distributed map of these communication-constrained environments that in addition exhibit perceptual aliasing due to repetitive structure. Since each robot operates with respect to a local reference frame, robots estimate relative transforms to other robots’ reference frames using observed environment correspondences in order to construct a consistent distributed map. Perceptual aliasing, or the incorrect association of observations from different areas in the environment, complicates the estimation of relative transforms between robots. In this work, we extend a robust distributed mapping formulation to operate using 3D sensors under hardware network limitations for operation in the communication constrained, repetitive structure of an underground mine.

Real-world communication constraints limit a robot from sharing large numbers of observations at high fidelity. Na¨ıvely simplifying sensor information leads to loss of unique features and an increase in perceptual aliasing. Towards sharing the most relevant subset of information, we develop a scan utility function based on information theoretic measures to assess a scan’s ability to reduce map uncertainty and featurebased place recognition approaches to assess a scan’s potential for containing shared observations between robots. Using the utility function to rank scans, we formulate an offer-response-request framework, Communication Constrained Information Routing (CCIR), that ensures operation under stringent bandwidth restrictions. In simulation results, CCIR decreases the required network usage for distributed mapping to 20.7% of a fixed-rate down-selection approach.

Given the ability to share rich 3D information over constrained networks, we pursue full 3D mapping via extensions to existing approaches including robustification techniques. The robust measures we introduce allow operation in the targeted mine environment even given substantial perceptual aliasing with outliers accounting for 98.1% of all detections. Furthermore, the developed CCIR framework allows robots to develop relative transforms while respecting network bandwidth constraints. Similar performance when operating using a fixed-rate down-selection approach over the same mine environment requires 7.69 times more data transmission.

Additionally, to enable operation in environments that exhibit perceptual aliasing that exceeds the performance characteristics of the developed CCIR framework, this thesis details first results for an approach that moves away from feature-based techniques and introduces a methodology utilizing Hierarchical Gaussian Mixture Models. Through regeneration of the point cloud from the HGMM model and Generalized Iterative Closest Point algorithms, we show that we are able to detect multi-robot loop closures accurately with an outlier rate 34% of that of feature-based methods.