A Multimodal 3D Perception-Based Autonomous Robotic Architecture for Underground Mining: Predictive Planning and Dynamic Risk-Aware Control

Segundo Manayay J.L.; Garay Yovera J.Y.; Chávez Chávez R.G.; Rodríguez Huiman F.; Castro Suazo F.A.; Antayhua Alvarez R.N.; Espinoza Coronel J.A. · Proceedings of the LACCEI international Multi-conference for Engineering, Education and Technology (2026) · DOI: 10.18687/laccei2026.1.1.2664

Resumen

Underground mining inspection remains hazardous, with rockfalls and structural collapses accounting for over 60% of fatal accidents in Latin America. This paper presents an autonomous quadruped architecture integrating LiDAR-inertial SLAM, risk-aware A* planning, a 12-DoF computed-torque controller, and on-device TinyML acoustic classification within a ROS 2 framework. Validation is conducted via high-fidelity simulation and hardware-in-the-loop experiments in a laboratory gallery mockup that reproduces the geometric and sensory constraints of underground tunnels. The central novelty lies in a cost function that dynamically incorporates terrain traversability and acoustic hazard probabilities-obtained from a 1D-CNN deployed on an Arduino Nano 33 BLE Sense-directly into real-time navigation, enabling sub-40 ms replanning. Locomotion control maintains center-of-mass deviation below 3 cm and joint tracking error of 3.2◦ RMSE over irregular terrain, while MQTT telemetry achieves sub-50 ms latency under emulated communication loss. Experimental results show 91.8% acoustic classification accuracy with <25 ms edge inference, an 87% reduction in hazard exposure compared to distance-only A*, and 96% obstacle traversal success. These findings demonstrate that a low-cost, edge-AI-enhanced quadruped can autonomously navigate mining-representative galleries, detect geomechanical and gas hazards in real time, and alert supervisors-establishing a foundation for reducing human entry into dangerous underground environments.

Instituciones: Universidad Nacional de Ingeniería