Large model-driven semantic SLAM for path planning in complex environments

Abstract

Addressing the extremely complex challenges of high-dynamic and highly reflective spatial perception in industrial environments, this paper innovatively proposes a large-model-driven vision-language cross-modal semantic SLAM and heuristic topology planning architecture. The system achieves zero-sample 3D geometric matching of open vocabularies through a deep dual-stream feature network and designs a semantic confidence-weighted core optimization mechanism to effectively suppress multiple environmental outlier artifacts. This successfully reduces the global absolute trajectory error limit in harsh factory environments to 3.12 cm, improving the system’s robustness against disturbances by nearly 87%. Leveraging a cloud-edge collaborative computing architecture, the model maintains 32 FPS high-frequency, high-fidelity odometry concurrent processing while deeply embedding common-sense prior logic into the underlying risk planning cost function. This enables large embodied robots to achieve a dynamic avoidance success rate of 96.4% under extreme spatial constraints, with the system’s global computation latency stabilized at 185 milliseconds, demonstrating significant potential for practical application.