Source-linked AI summary
Autonomous Teamed Exploration of Subterranean Environments using Legged and Aerial Robots
Mihir Kulkarni, Mihir Dharmadhikari, Marco Tranzatto, Samuel Zimmermann, Victor Reijgwart, Paolo De Petris, Huan Nguyen, Nikhil Khedekar, Christos Papachristos, Lionel Ott, Roland Siegwart, Marco Hutter, Kostas Alexis
TL;DR
Subterranean environments combine complex, large-scale geometries with challenging communications, making teamed autonomous exploration demanding. The paper pairs a single-robot planner for legged and aerial robots with a framework that coordinates exploration through shared maps and global frontiers. It evaluates the strategy in a Swiss underground mine and a three-robot simulation.
Problem
Subterranean exploration must handle complex, large-scale geometries and challenging communications while coordinating heterogeneous robots.
Source: §I · p. 1Method
GBPlanner2 supports individual legged and aerial robot exploration, while COHORT uses aligned submaps to identify global frontiers and guide team repositioning.
Source: §I · p. 1, A. Graph-based Subterranean… · p. 3, B. Cooperation for Heterogeneous Robot… · p. 4Results
Takeaways & Limitations
The framework combines individual exploration autonomy with map-based coordination and is described as robust against communications-deprived environments.
Source: §VI · p. 6
Abstract
from arXiv · showhide
This paper presents a novel strategy for autonomous teamed exploration of subterranean environments using legged and aerial robots. Tailored to the fact that subterranean settings, such as cave networks and underground mines, often involve complex, large-scale and multi-branched topologies, while wireless communication within them can be particularly challenging, this work is structured around the synergy of an onboard exploration path planner that allows for resilient long-term autonomy, and a multi-robot coordination framework. The onboard path planner is unified across legged and flying robots and enables navigation in environments with steep slopes, and diverse geometries. When a communication link is available, each robot of the team shares submaps to a centralized location where a multi-robot coordination framework identifies global frontiers of the exploration space to inform each system about where it should re-position to best continue its mission. The strategy is verified through a field deployment inside an underground mine in Switzerland using a legged and a flying robot collectively exploring for 45 min, as well as a longer simulation study with three systems.
I. INTRODUCTION
The paper addresses the difficulty of autonomous exploration in demanding subterranean environments through two synergistic methods for legged and aerial robots. It evaluates the strategy in a Swiss underground mine and in large-scale simulation.
I. INTRODUCTION
- Subterranean environments remain especially demanding for robot navigation and autonomy, despite progress in robotic exploration across other mission settings.The paper cites mines, subway infrastructure, and cave networks as examples.Source: §I · p. 1
- The work combines GBPlanner2 for individual robot exploration with COHORT for coordinating heterogeneous teams in large, branching subterranean environments.The methods are intended to support coordinated exploration while each robot retains individual exploratory capacity.Source: §I · p. 1
- Evaluation includes a Hagerbach underground-mine deployment with legged and aerial robots exploring for 37 min and 8 min, respectively, plus a large-scale simulation study.The authors report efficient autonomous exploration resilient to environmental scale, challenging terrain and geometry, and difficult communications.Source: §I · p. 1
II. RELATED WORK
Prior work spans single-robot planning, multi-robot coordination, and subterranean exploration, including methods for legged and flying systems. This paper contributes single-robot planning and team coordination tailored to subterranean deployment.
II. RELATED WORK
- Prior single-robot exploration includes next-best-view sampling, frontier detection, graph and tree planning, receding-horizon methods, optimization, information-theoretic approaches, learning, and localization-drift-aware planning.The passage also surveys multi-robot methods using grid maps, distributed inference, volumetric gain, heterogeneous teams, and limited-communication or endurance considerations.Source: §II · p. 2
- Subterranean exploration research includes methods for quadrupeds, aerial platforms, legged and flying systems, hierarchical mapping, mine and cave topologies, and field robotics.The paper situates its contribution in this growing area of work.Source: §II · p. 2
- The paper contributes coordination and single-robot planning for legged and flying teams, targeting resilient autonomy, challenging terrain, and diverse large-scale geometries.The related-work passage states that these capabilities address reliable single-robot operation when communication is unavailable.Source: §II · p. 2
III. PROBLEM FORMULATION
The paper formulates exploration as mapping a bounded subterranean volume with a heterogeneous team subject to sensor and motion constraints. Robots retain substantial individual autonomy because subterranean topology and materials may prevent continuous connectivity.
III. PROBLEM FORMULATION
- The environment is represented by an occupancy map of free, occupied, or unknown voxels, and some hollow, narrow, or occluded sections may be unobservable.The residual volume consists of voxels that cannot be perceived from any collision-free configuration.Source: §III · p. 2
- The team must find collision-free paths that identify free and occupied parts of a bounded volume, leaving only residual space that no collision-free configuration can perceive.Feasible paths must respect sensor field of view and range as well as robot motion constraints.Source: §III · p. 2
- Subterranean topology and materials may prohibit continuous network connectivity, so multi-robot exploration relies on substantial individual autonomy.The formulation allocates a volume to a robot at a given time and for a specified duration.Source: §III · p. 2
- Legged robots must traverse supportive ground and remain within feasible inclination limits, while paths must also respect sensor and motion constraints.The environments motivating the formulation include branches, multiple levels, large rooms, vertical structures, steep slopes, and anomalous terrain.Source: §III · p. 2
IV. PROPOSED APPROACH
The proposed approach pairs a comprehensive single-robot exploration planner with a coordination method for ground and flying teams. Its planner is revised for subterranean exploration across aerial and legged robots.
IV. PROPOSED APPROACH
- The approach combines increased individual robot autonomy through a comprehensive exploration planner with multi-robot ground and flying exploration.The two methods are GBPlanner 2.0 and COHORT.Source: §IV · p. 2
- The revised planner's functional steps are outlined for aerial and ground (legged) robots.The figure caption identifies the planner as graph-based and subterranean.Source: Figure 2 · p. 3
A. Graph-based Subterranean Exploration Revisited
GBPlanner2 combines local exploration planning with global repositioning for single-robot exploration of unmapped underground volumes. Its local graph construction and gain calculation are adapted to varied terrain and computational constraints.
A. Graph-based Subterranean Exploration Revisited
- GBPlanner2 retains local/global planning: the local planner finds motion- and perception-feasible exploration paths, while the global planner repositions the robot when local gain is insufficient or homing is needed.Global repositioning is triggered when the local layer cannot find a path of significant exploration gain or to ensure return within endurance limits.Source: A. Graph-based Subterranean… · p. 3
- Local graph bounds adapt to explored geometry, and ground-robot edges are checked for mapped support and feasible inclination to handle slopes, vertical obstacles, and ditches.Flying-robot edges are collision-free straight-line segments; ground-robot validation also checks nearby support voxels to address incorrectly mapped free ground.Source: A. Graph-based Subterranean… · p. 3
- Ground-free “hanging” vertices can contribute to volumetric gain, but their gain is scaled by e^-γH and their edges cannot be commanded to the robot.This distinguishes vertices useful for estimating visible volume from traversable destinations.Source: A. Graph-based Subterranean… · p. 3
- Dijkstra paths provide candidate routes whose volumetric gain is accumulated over vertices; GBPlanner2 can approximate gain using clustered leaf vertices to reduce computation.Nearby vertices within radius ρ can reuse a calculated gain, a trade-off aimed at computationally constrained micro flying robots in large environments.Source: A. Graph-based Subterranean… · pp. 3–4
B. Cooperation for Heterogeneous Robot Teams
COHORT coordinates robots that explore independently with GBPlanner2 by integrating submaps into a shared map and directing robots toward selected global frontiers. The process accommodates intermittent communication by using previously recorded communication locations for backtracking.
B. Cooperation for Heterogeneous Robot Teams
- Robots send selective, compressed submap data to a centralized hub, which aligns map estimates; this requires a communication link that may be intermittent.All robot poses and maps must share an inertial frame for COHORT operation.Source: B. Cooperation for Heterogeneous Robot… · p. 4
- Submaps represent bounded map windows, started anew when either the 45 s time slot or 20 m displacement threshold is exceeded.The windows use a volumetric representation developed with Voxblox.Source: B. Cooperation for Heterogeneous Robot… · p. 4
- COHORT finds submap frontiers by wavefront propagation over unvisited voxels neighboring visited ones, then clusters neighboring frontier points.The resulting frontier points are centroids of unknown voxels neighboring free mapped voxels.Source: B. Cooperation for Heterogeneous Robot… · p. 4
- Each submap contributes a collision-free subgraph to COHORT’s global graph, which grows as robots’ explored regions are integrated.COHORT samples free points within the submap and connects them within a radius, including sampled points along the robot trajectory.Source: B. Cooperation for Heterogeneous Robot… · p. 4, Figure 3 · p. 5
- COHORT periodically selects a promising global frontier near the requesting robot, routes it through the merged graph, and updates its GBPlanner2 goal volume around that frontier.Candidate frontiers are re-clustered across submaps, ranked, and filtered to the best nC%; the nearest among those to the requesting robot is selected.Source: B. Cooperation for Heterogeneous Robot… · p. 5
- If communication is unavailable, a robot continues exploring briefly when its assigned volume remains incomplete, then backtracks to a previous communication point.If the assigned volume is complete, it returns to the last point where communication was available, iterating to earlier points if needed.Source: B. Cooperation for Heterogeneous Robot… · p. 5
V. EVALUATION STUDIES
The evaluation includes a mine deployment with a legged and a flying robot, alongside a three-robot simulation of teamed exploration. In the deployment, both systems explored autonomously and returned home; the simulation includes operation beyond communication range.
V. EVALUATION STUDIES
- 37 min for ANYmal C and 8 min for RMF-Obelix: both explored the Hagerbach mine without manual interruption, repositioned autonomously, and returned home.ANYmal C navigated slopes; RMF-Obelix was a 1.4 kg aerial robot, and the two systems used different processors and planner settings.Source: §V · p. 5, Figure 4 · p. 6
- The mine deployment used ANYmal C with a Velodyne VLP-16 and RMF-Obelix with an OUSTER OS0-64 LiDAR; Table I summarizes their GBPlanner2 parameters.The systems also differed in processing capabilities and total endurance, mission time, average processing time, and mapping resolution.Source: §V · p. 5, Table I · p. 6
- The simulation evaluates COHORT with three flying robots and models communication constraints, including exploration beyond communications range and backtracking after an additional 150 s.COHORT repositioned robots toward global frontiers in the commonly explored map while GBPlanner2 continued onboard exploration.Source: §V · p. 6
VI. CONCLUSIONS
The conclusion presents a framework pairing resilient individual-robot exploration planning with multi-robot coordination over aligned maps to identify global frontiers, including in communications-deprived environments. It reports large-scale field deployments and simulations as evaluation.
VI. CONCLUSIONS
- The framework combines resilient exploration path planning on individual robots with multi-robot coordination for complex subterranean environments.The conclusion describes this as a synergy intended for large-scale, multi-branching settings with challenging terrain.Source: §VI · p. 6
- The coordination framework uses aligned maps to identify global frontiers and is described as robust in communications-deprived environments.The figure shows three flying robots exploring beyond communication range and returning after reaching a frontier.Source: §VI · p. 6, Figure 5 · p. 6
- The authors evaluate the framework through large-scale field deployments and simulations.Source: §VI · p. 6