A Bioinspired Neuromorphic Chip Based on Protein-Protein Interaction Network Topology and Its Application in Adaptive Gait Control for Legged Robots

Abstract

Traditional footstep controller chips of the legged robot have problems such as low energy utilization rate, lack of flexibility, and poor environmental resistance ability. In order to solve this problem, this study proposes a neuromorphic chip design based on PPI network topology, which mimics the small-world property, scale-free degree distribution, and modular structure observed for the PPI networks using a set of protein network processors (PNPs).In memristor crossbar arrays, synaptic weights can be stored and computed in situ to eliminate the memory-processor bottleneck. A  three-level motion hierarchy deals with control on different timescales: a reflex level for millisecondscale posture stabilization, a rhythm level to generate periodic gait patterns, and a strategy level for long-horizon motion planning and terrain adaptation. Fabricated in 28 nm CMOS, the chip contains 1024 PNPUs and draws a total power of 2.5W. Experiments with a quadruped robot show that our chip is able to achieve better than 10x energy efficiency over traditional model predictive control running at 1.5 m/s walkspeed on flat ground, with terrain transitions completed within 0.8s. With a quarter of processing units disabled, we still retain more than 80% of locomotion capability, verifying that our model generalizes to unseen situations.