The Perception, Decision-Making, and Execution of the Centipede-Like Rescue Robot

This paper explores the challenge of closing the loop on robotic systems inspired by centipedes in disaster zones using various technologies. To achieve this, it presents a cognitive architecture that employs a Large Language Model (LLM) combined with a Lang Graph to connect higher-level thought with lower-level controlling devices. This cognitive architecture allows us to utilise different types of sensors to create a hierarchy of actions using real-time data to coordinate multi-agent activities. It evaluates three different approaches: lightweight edge closed-loop, RL-driven, and large model collaborative-hubbing, and discusses different technical merits in terms of speed, accuracy, and power consumption. It addresses fundamental limitations found with perceptions, communications, and cognitive processes. Additionally, it describes solutions such as adaptive sensor fusion, federated learning, and hierarchical decision-making. It has tested the framework’s effectiveness in both field drills and collaborative research projects. Finally, it discusses potential advances in rescue robots toward developing greater autonomy through future research on developing hybrid systems that incorporate both neuro-symbolic and embodied active perception.