A Hierarchical Perception-Action Framework for Adaptive Legged Locomotion in Unstructured Environments

Reliable legged locomotion in unstructured environments requires the integration of fast local response, terrain-aware gait adaptation, and longer-horizon navigation. In many existing systems, these functions are implemented in separate modules, which weakens real-time interaction between low-level control and high-level planning. This paper presents a Hierarchical Perception-Action Coupling (HiPAC) framework organized as a three-level control hierarchy supported by a sensory layer embedded in the onboard robot controller, operating at different time scales. The framework comprises a reactive layer for immediate terrain response and gait execution, a context layer for local environmental recognition and locomotion modulation, and a planning layer for cognitive-map construction, path generation, and replanning. The sensory layer acquires exteroceptive and proprioceptive measurements and forwards them to the higher layers through a lightweight topological representation. In HiPAC, the layers exchange compact topological representations instead of dense raw sensory data, which keeps memory and computation bounded. The framework is evaluated in physics-based simulation on a quadruped robot across flat, rough, and multi-story environments, and is further demonstrated in real-robot deployment. The results show that the reactive layer maintains response times below 30ms, the proposed topological attention mechanism reduces processing cost by up to 70% relative to uniform-density representations, and local environment reconstruction remains below 18kB during extended exploration. These findings demonstrate that HiPAC enables efficient integration of reactive control, terrain adaptation, and navigation under strict computational constraints.