RISC-V Summit

Advanced computing platforms for autonomous drone navigation are extremely challenging drivers of Internet-of-Things edge intelligence capabilities due to the mix of requirements in terms of security, safety and high performance. We present a fully RISC-V-based SoC for high-performance and secure autonomous drone navigation. The SoC includes a dual-core CVA6 core with full support for Linux OS managing the main system functions and an RTOS such as Nuttx for time-critical tasks. The SoC is supported by a complete set of peripherals, and a parallel accelerator based on the open-source PULP platform and a root-of-trust subsystem based on the open titan project. As a capability demonstrator, we present a human pose estimation task based on a novel CNN. Inference runs in real-time, up to 135 frame/s, consuming less than 87 mW on a COTS PULP-based SoC aboard a 27-grams nano-UAV, where it achieves 1 GOPS with an efficiency of more than 12 GOPS/W. The in-field deployment of our visual NN demonstrates a closed-loop top-notch autonomous navigation capability. Our novel RISC-V-based SoC is expected to push this cutting-edge end-to performance even further, as much as 2.6 GOPS and 92 GOPS/W.