TL;DR
The Luojia Explorer is a robust multi-robot system designed to autonomously map unknown environments for digital twin construction. By integrating centralized collaborative SLAM (C-SLAM) with a distributed "Explorer-Collector" planning strategy, the system achieved a high exploration efficiency of 38.15 m³/s in real-world campus trials.
What Problem Does This Paper Address?
Creating "digital twins"—precise 3D virtual models of physical spaces—is essential for planetary exploration, underground mining, and urban planning. While single-robot systems can map these areas, they are often too slow for large-scale environments or time-critical missions.
Multi-robot systems (MRS) offer a solution through parallelization, but they introduce a "coordination tax." Researchers face three major hurdles:
- Localization: How do you merge individual robot maps into a single, consistent global frame?
- Task Allocation: How do you prevent robots from exploring the same area twice?
- Communication: How do you ensure reliable data exchange in complex environments where signals might drop?
Methods
The authors introduce the Luojia Explorer, a hardware-software framework that uses a team of heterogeneous robots (one wheeled, one tracked) to solve these issues.
The Algorithmic Stack
- Centralized C-SLAM: Each robot runs its own local LiDAR-Inertial Odometry (based on the Fast-LIO2 algorithm). The system designates the most computationally efficient robot as the "Master Node." This node collects local data from all peers and performs global pose graph optimization to align everyone into a unified coordinate system.
- Distributed Collaborative Planning (DCARE): While SLAM is centralized, planning is distributed. Each robot analyzes a global 3D voxel map to identify "frontiers"—the borders between known and unknown space.
- Explorer vs. Collector Modes: This is a standout feature. Robots normally operate in Explorer mode, prioritizing large, open areas for rapid coverage. However, if a robot detects too many small, neglected frontiers (which might be narrow corridors or doors), it switches to Collector mode. In this mode, it solves a "Traveling Salesman Problem" (TSP) to efficiently visit all small gaps before returning to high-speed exploration.
Hardware & Communication
The robots use Livox Mid360 LiDARs and communicate via a 1.5G wireless Mesh network. To bridge the gap between legacy stability and modern flexibility, the system uses a ROS1-to-ROS2 bridge, leveraging ROS2’s Data Distribution Service (DDS) to prioritize critical data packets even during network congestion.
Datasets
The system was not tested on standard simulated benchmarks; instead, the authors conducted real-world experiments in a complex campus environment. The test area included buildings, narrow passages, vehicles, and varied terrain to simulate the "unknown spaces" the system is designed for.
Key Insights and Ideas
- Heterogeneity as a Strength: By using different chassis types (differential drive vs. tracked), the system is more adaptable to varied terrains found in unknown spaces.
- The "Aha" of Mode Switching: Purely greedy exploration often misses small openings that lead to massive unexplored rooms. The "Collector" mode ensures these "bottleneck" frontiers are handled without sacrificing the speed of the "Explorer" mode.
- Information Gain Function: The target selection isn't just about distance. The system uses a complex gain function that weights volume gain, path length, direction changes (to avoid "jittery" movement), and the positions of teammates (to ensure dispersion).
Key Results
The real-world deployment of two robots yielded impressive quantitative metrics:
- Exploration Efficiency: 38.15 m³/s.
- Total Volume Mapped: 9557.09 m³ in just over 4 minutes (250.45 seconds).
- Path Quality: The robots maintained an average path length of 66.74 meters, indicating high-quality, non-redundant trajectories.
- Map Consistency: Qualitative results showed a seamless global point cloud with no "layering" or ghosting, proving the C-SLAM was effective at merging the robots' perspectives.
Limitations and Future Work
While the system is highly effective, the authors identify communication architecture as the primary area for improvement. In extremely shielded environments (like deep mines), even a Mesh network might struggle. Future iterations will likely focus on optimizing how much data is sent over the wire to maintain coordination in "low-bandwidth" or "intermittent-link" scenarios. Additionally, scaling the system from two robots to a larger swarm will require further stress-testing of the centralized Master Node's processing limits.
Summary updated