RoboNav-Arm uses agentic AI for real-time obstacle avoidance in cluttered environments

RoboNav-Arm, presented by Aachal Sharma and Narendra Kumar Dhar on arXiv, is an agentic AI-driven navigation and obstacle avoidance system designed for robotic manipulators operating in unstructured, cluttered environments. Traditional methods depend on prior knowledge or fixed perception pipelines, limiting adaptability. RoboNav-Arm replaces this with a modular framework that performs real-time obstacle detection, 3D localization, and ground surface geometry estimation. It then generates a structured semantic report detailing obstacle positions, shapes, and whether they lie inside, outside, or within critical interaction zones. A central coordination module manages tool invocation (e.g., memory and MoveIt collision scene updates), facilitates inter-module communication, and continuously monitors task progress until completion. The planning module dynamically selects the most suitable motion planning algorithm—RRTConnect, RRT*, or BiTRRT—based on the current environment and goal requirements. The resulting trajectory is analyzed and refined to guarantee safe, collision-free execution. The system was validated in Gazebo Classic, demonstrating robust performance in dynamic scenarios where obstacles appear unpredictably. By leveraging agentic AI to perceive, decide, and act autonomously, RoboNav-Arm represents a significant step toward truly adaptable industrial and service robots that can operate safely alongside humans or in cluttered warehouses.