Rapid and High-Fidelity Subsurface Exploration with Multiple Aerial Robots

TL;DR

This work addresses rapid, high-detail subsurface exploration using a multi-robot aerial team under severe communication constraints. It introduces a Gaussian Mixture Model (GMM) based distributed mapping framework, where keyframe GMMs are exchanged and integrated across robots via coordinate transforms so that occupancy grids remain accurate yet compact. Planning is driven by information-theoretic receding-horizon control using Monte Carlo Tree Search to maximize a CSQMI objective, with practical optimizations to maintain real-time performance and safety. The results show significant memory efficiency and faster exploration than OG/OM baselines, validated by both hardware experiments in a cave and constrained-bandwidth simulations, indicating strong potential for planetary subsurface missions and other bandwidth-limited environments.

Abstract

This paper develops a communication-efficient distributed mapping approach for rapid exploration of a cave by a multi-robot team. Subsurface planetary exploration is an unsolved problem challenged by communication, power, and compute constraints. Prior works have addressed the problems of rapid exploration and leveraging multiple systems to increase exploration rate; however, communication considerations have been left largely unaddressed. This paper bridges this gap in the state of the art by developing distributed perceptual modeling that enables high-fidelity mapping while remaining amenable to low-bandwidth communication channels. The approach yields significant gains in exploration rate for multi-robot teams as compared to state-of-the-art approaches. The work is evaluated through simulation studies and hardware experiments in a wild cave in West Virginia.

Figures (6)

• Figure 1: Cave exploration with two aerial robots in West Virginia, USA. A video of the flight can be accessed at the following link: https://youtu.be/osko8EKKZUM.
• Figure 2: (Left) Overview of the rapid multi-robot exploration framework and (Right) aerial systems used in experiments in this work.
• Figure 3: Overview of the distributed mapping approach. \ref{['sfig:data']} Robot $i$ shown in red, takes a sensor observation shown in colors varying from red to purple and \ref{['sfig:gmm_sent']} learns a GMM (shown in red). If the GMM is determined to be a keyframe both the GMM and sensor pose are transmitted to robot $j$ (shown in green). \ref{['sfig:map_update']} The GMM and the sensor pose are transformed into the frame of robot $j$ and used to update the occupancy.
• Figure 4: Fidelity and memory usage evaluation of several mapping approaches. \ref{['sfig:crevice_rgb']} and \ref{['sfig:crevice_depth']} illustrate data from a representative environment the robot may encounter in the cave. A potential passage is circled in cyan. \ref{['stab:memory_table']} highlights significant reduction in memory usage required by the GMM approach as compared to the OG and OM approaches. \ref{['sfig:crevice_gmm']} Resampled points from the GMM are shown in red. \ref{['sfig:crevice_octomap_0.025m']}--\ref{['sfig:crevice_octomap_0.1m']} illustrate the OctoMap representation with leaf sizes varying from 0.025m to 0.1m. Leaf voxels are shown in red and larger voxels in yellow.
• Figure 5: Rapid and communication efficient exploration of a cave with a team of two aerial robots. \ref{['sfig:env']} illustrates the environment with the two robots (R1 and R2) and the WiFi router used for communication. \ref{['sfig:gmm_cave']} illustrates the final GMM maps generated on the base-station. \ref{['sfig:speeds_cave']} shows the percentage density plots for linear speeds and yaw rates as measured by the visual-inertial navigation system during flight. \ref{['stab:rapps_comm']} highlights that the GMM approach requires significantly less memory to represent the combined map as compared to state-of-the-art approaches. In the context of transmitting this data using a channel with capacity $0.25M\bit\per s$, it would take significantly less time for the GMM approach as compared to the other approaches. A video of the flight can be accessed here: https://youtu.be/osko8EKKZUM.