Safe Operation in Clean Energy: A Knowledge-Enhanced Generative Framework for Insulated Tool Acquisition

Abstract The rapid transition towards unattended operation and maintenance (O&M) in large-scale clean energy infrastructures urgently requires autonomous robots capable of performing active physical interventions. However, executing tool acquisition in these hazard-critical environments poses severe multifaceted risks, including the risk of DC arcing and thermal exposure. Traditional robotic grasping methods typically prioritize geometric stability but lack the semantic awareness required to identify and avoid hazardous functional ends, leading to a high risk of unsafe interactions and compromised downstream usability. To overcome this critical bottleneck, we propose Knowledge-Driven Safety Grasping (KDSG), a novel framework designed to improve hazard-aware operational safety during autonomous tool acquisition. KDSG integrates two core modules: Safety-Aware Grounding and Enrichment (SAGE) and Physically Constrained Generative Grasping (PCGG). SAGE leverages a domain-specific Hazard Knowledge Graph coupled with Vision-Language Models (VLMs) to semantically truncate hazardous parts from the perceptual input, explicitly isolating safe interaction regions. Building upon this safe geometry, PCGG utilizes a distance-based generative representation to predict geometrically conformal grasps, followed by a strict Physical Safety Filter that enforces deterministic collision avoidance. We evaluate KDSG in a high-fidelity physics simulator using a multi-fingered dexterous hand. Extensive comparative and ablation experiments demonstrate that our framework reduces the Safety Violation Rate (SVR) to 0.0% under our simulator-defined metric, while maintaining a robust Operational Safety Success Rate (OSSR) of 70.3%. This work provides a promising framework for the deployment of intelligent robotic systems in hazard-critical clean energy infrastructures.