Edge-Aided Multi-Modal Collaborative SLAM for Resource-Constrained Underground Robots

Simultaneous localization and mapping (SLAM) of coal mine robots (CMRs) have become inevitable for monitoring, rescue, and transportation in coal mines. However, low illumination and single structures inevitably lead to sensors degradation in mine tunnels. Multi-modal fusion SLAM offers sensing complements but aggravates the consumption of computing resources, which is a huge challenge for CMRs with resource constrained memory and processors. In this paper, we design and deploy an edge-aided multi-modal collaborative SLAM algorithm (EM-CoSLAM) with LiDAR-inertial-visual fusion for constrained CMRs. EM-CoSLAM provides real-time mobile LiDAR tracking and precise edge visual optimization in parallel by decoupling and restructuring task modules. To reduce end-to-end latency and enhance system performance, EM-CoSLAM devises an adaptive offloading strategy to determine the optimal image frame by minimizing the pose uncertainty and further proposes a fusion pose graph optimization method. We completely implement EM-CoSLAM on the CMR and extensively evaluate it in various datasets and scenarios. The results indicate that EM-CoSLAM can achieve 6.7cm and 16fps performance on the CMR, outperforming existing solutions by more than 14%.