Centralized and decentralized control for multi-legged locomotion on rough terrains

Centipedes traverse rough, complex ground with ease, whereas multi-legged robots often struggle to achieve comparable performance. This gap mainly stems from limitations in their control architecture. Previous studies have explored decentralized control (DC) based on local central-pattern generators (CPGs) (e.g., Ijspeert et al., JEB, 2023), which perform well on specific terrains but fail in more complex environments. We hypothesize that integrating centralized control (CC) with DC is crucial for the intelligent and versatile locomotion observed in centipedes and aim to clarify (1) what is CC in centipedes, (2) how to coordinate CC and DC, and (3) how to characterize their coordination. Building on the DC-based centipede locomotion simulation in the FARMS framework by Prof. Ijspeert’s group, we tested “seed gaits” generating different numbers of body waves and observed that more waves enhance robustness on rough terrain but reduce speed on flat ground. Motivated by this, we introduce a centralized “mode selector” that adapts the seed gait of DC based on global sensory feedback. Frequency sweeps reveal that low-frequency CC decisions already achieve near-maximal performance, whereas effective DC requires high-frequency feedback, highlighting a correlation between feedback timescale and information concentration. Together, this hierarchical architecture advances our understanding of multi-legged locomotion in both animals and robots.