A Unified Microrobotic Visual-Perception Processor with 62.2-FPS/mm² and 103-uJ/frame Navigation in 28nm

Microrobotic technology has made significant advancements recently, while the payload and power limitations restrict its development toward greater intelligence, for example, the constraint of $\sim 100 \mathrm{mg}$ and $\sim 100 \mathrm{~mW}$ in an insect-sized robot [1]. As an essential role in robotic navigation, visual perception usually has two critical components: localization and neural-network (NN) inference. The former has unique operators and dataflow featuring small-kernel filtering in the vision front-end (VFE) and, more importantly, the linear solver in the optimization back-end involves heavy Cholesky decomposition which is hard to cast into matrix-vector multiplication (MVM) [2]. These distinct dataflows usually lead to heterogeneous systems with separated fixed-function accelerators [3–7]. The visualinertial odometry in [3] focuses on localization and exploits fixedpattern sparsity in the back-end, which tradeoffs efficiency with restrictions on applications’ data distributions. The visual-perception SoCs in [5, 6] handle localization front-end and NN inference in robots efficiently, with the optimization back-end left external. However, severe system throughput degradation occurs when the computedemand ratio deviates from the native supply of the accelerators in diverse application scenarios [8], as shown in Fig. 1.