Three Robots Advance Decentralized C-Slam for Planetary Exploration and Mapping

Scientists are tackling the complex problem of enabling multi-robot missions in environments where pre-existing maps and communication networks are unavailable, a crucial capability for exploring destinations like the Moon and Mars. Pierre-Yves Lajoie, Karthik Soma, and Haechan Mark Bong, all from Polytechnique Montreal, along with Lemieux-Bourque, Zhang, Varadharajan et al, present a detailed analysis of decentralized collaborative simultaneous localization and mapping (C-SLAM) conducted with three robots in a planetary analogue terrain. Their research is significant because it highlights the impact of limited and intermittent communication on C-SLAM performance, alongside the specific localization difficulties presented by planetary landscapes, and importantly, they publicly release a novel dataset of real-time inter-robot communication metrics to facilitate further investigation in this challenging field?

This research addresses the significant challenge of enabling robot teams to operate autonomously in unknown environments, such as the Moon, Mars, and beyond, without reliance on pre-existing infrastructure. The team achieved this breakthrough through extensive experiments involving three robots navigating a Mars analogue terrain and communicating via an ad-hoc network, meticulously examining the impact of limited and intermittent communication on C-SLAM performance. A key innovation of this work is the introduction of a new dataset, comprising real-time peer-to-peer inter-robot throughput and latency measurements, designed to facilitate future research into communication-constrained, decentralized multi-robot operations.

The study meticulously evaluated the performance of Swarm-SLAM, a decentralized C-SLAM framework, in a challenging planetary analogue setting. Experiments were conducted at the Canadian Space Agency Mars Yard, where the robots simultaneously explored the simulated Martian terrain, collaborating through peer-to-peer communication to build a map and determine their positions. This approach contrasts with traditional centralized C-SLAM systems, which are vulnerable to network disruptions and single points of failure, making them less suitable for the harsh realities of space exploration. The researchers focused on building a robust and resilient system capable of operating with frequent and prolonged disconnections, a critical requirement for missions where communication with Earth is limited or unreliable.
This work establishes the importance of decentralized techniques for C-SLAM, purposefully designed for ad-hoc networking and resilience against prolonged disconnections. The team’s novel dataset includes not only standard sensor data like LiDAR and IMU readings, but also crucial real-time measurements of inter-robot communication performance, a feature absent in previous datasets. This detailed communication data allows for a more accurate assessment of the resource consumption of C-SLAM algorithms and facilitates the development of more efficient multi-robot systems. The research highlights the unique localization challenges presented by planetary analogue environments, including vibrations, a lack of distinctive features, and perceptual aliasing, all of which impact the accuracy and efficiency of C-SLAM.

Furthermore, the study provides a thorough analysis of the accuracy and efficiency of decentralized C-SLAM, revealing limitations in current approaches and identifying key areas for future research. The experiments, conducted with a three-robot system connected through an ad-hoc network, demonstrate the feasibility of this approach in a realistic planetary analogue environment. The resulting dataset, available on IEEE DataPort, is intended to serve as a valuable resource for the space robotics and C-SLAM research communities, fostering further innovation in autonomous multi-robot exploration. The work opens avenues for developing more robust and efficient robotic teams capable of tackling the challenges of exploring distant worlds.

Multi-robot C-SLAM with limited communication and LiDAR

Scientists conducted a collaborative simultaneous localization and mapping (C-SLAM) study employing three robots within an analogue planetary terrain to assess the impact of limited communication on multirobot operations. The research team engineered a system where robots exchanged 3D keypoints and full scans to perform precise geometric registration, establishing accurate positional and rotational links between them. These resulting pose measurements, termed loop closures, were integrated into the robots’ pose graphs to merge maps and refine state estimation accuracy. When processing LiDAR scans, the team utilized robust methods like TEASER++ to achieve accurate registration even without initial pose estimations, a crucial capability given the robots’ lack of prior knowledge regarding relative positions.

The C-SLAM back-end tasked with estimating robot poses and maps leveraged pose graph optimization to reduce computational costs associated with large-scale SLAM. This method marginalizes environmental features into inter-pose measurements, focusing solely on pose optimization. Rather than implementing a distributed solver, the study pioneered a streamlined approach by dynamically electing a single robot to perform computations during each rendezvous, mitigating the bookkeeping and iterative demands of distributed systems. Researchers addressed the challenge of perceptual aliasing, where similar locations cause incorrect data associations, by integrating Pairwise Consistency Maximization (PCM) to identify the largest set of consistent inter-robot measurements.

To enhance robustness, the elected robots employed the Graduated Non-Convexity (GNC) algorithm for pose graph optimization, building upon previous work extending GNC for distributed implementations. Recognizing the limitations of vision-based loop closure detection in planetary analogue environments, the team prioritized LiDAR for mapping due to its independence from ambient light and precise 360° range measurements. Cameras were integrated with LiDAR to leverage complementary properties and improve loop closure detection. The study’s methodology also involved collecting a novel dataset containing real-time peer-to-peer inter-robot throughput and latency measurements, designed to support future research on communication-constrained, decentralized multirobot operations. This dataset, combined with the innovative optimization and sensor fusion techniques, enables further advancements in multirobot missions for environments like the Moon and other planets.

Swarm-SLAM performs robustly with limited communication between agents

Scientists conducted experiments to advance decentralized collaborative simultaneous localization and mapping (C-SLAM) technology, crucial for multirobot missions in environments lacking pre-existing infrastructure. The research focused on three robots operating within an analogue terrain, assessing the impact of limited and intermittent communication on C-SLAM performance and the unique localization difficulties presented by planetary-like surfaces. The team meticulously measured real-time peer-to-peer inter-robot throughput and latency, generating a novel dataset publicly available for future research on communication-constrained multirobot operations. Experiments revealed the Swarm-SLAM decentralized C-SLAM algorithm’s accuracy in challenging terrains, highlighting inherent localization issues in planetary analogue environments.

The study quantified the trade-offs between accuracy and resource efficiency, specifically examining the balance between communication and computing demands during calibration. Data shows that robust loop closure detection is a major challenge in the C-SLAM front-end, requiring efficient methods for detecting and computing inter-robot connections. Robots exchanged compact descriptors, such as image descriptors and LiDAR scan descriptors, to enable place recognition and identify overlaps in mapped environments. Measurements confirm that high similarity scores between descriptors indicated potential loop closure candidates, triggering the exchange of more detailed data like 3D keypoints or full scans for precise geometric registration.

The team employed robust methods like TEASER++ to ensure accurate registration of LiDAR point clouds, even without initial pose estimations, a critical capability for C-SLAM. The C-SLAM back-end was tasked with estimating the most likely robot poses and maps from noisy data, utilizing pose graph optimization to reduce computational costs. Results demonstrate that streamlining optimization by dynamically electing a single robot to perform computations during each rendezvous proved effective. The work identified that distributed multi-robot solvers, while improving scalability, often require extensive bookkeeping and can be hampered by network delays. This research delivers a valuable dataset, including real-time throughput measurements and latency figures, to support future advancements in communication-constrained, decentralized multirobot operations for space exploration and beyond.

Planetary C-SLAM under communication constraints presents unique challenges

Scientists have demonstrated the importance of decentralized collaborative simultaneous localization and mapping (C-SLAM) for multirobot missions in environments lacking pre-existing infrastructure. Their research involved experiments with three robots operating in an analogue planetary terrain, communicating via an ad hoc network, and focused on the impact of limited and intermittent communication on C-SLAM performance. A novel dataset was also introduced, containing real-time inter-robot communication measurements to facilitate further research in this area. The findings highlight significant challenges in deploying C-SLAM in planetary environments, particularly regarding difficult terrain and limited resources.

Experiments revealed that flat terrain and a lack of distinctive features can lead to inaccurate 3D registrations, with similar locations appearing indistinguishable despite being spatially separated. Communication bandwidth emerged as a primary constraint, with the C-SLAM front-end consuming the majority of available resources. The authors acknowledge a trade-off between communication and accuracy, and note that lossy data compression techniques, while potentially beneficial, require careful tuning to avoid discarding crucial features in sparse environments. Future research should focus on reducing the communication demands of C-SLAM front-ends through techniques like point cloud compression, descriptor-based representations, or voxelization. Improving accuracy without increasing communication and computational demands remains an open challenge, potentially addressed by refining algorithms, exploring new sensing paradigms, or developing more efficient data fusion techniques. The introduced dataset will support ongoing investigations into communication-constrained, decentralized multirobot operations and the development of more robust SLAM algorithms for challenging environments.