MetaE2RL: Toward Meta-Reasoning for Energy-Efficient Multigoal Reinforcement Learning With Squeezed-Edge You Only Look Once

Meta-reasoning shows promise in efficiently using the computational resources of tiny edge devices while performing highly computationally intensive reinforcement learning (RL) algorithms. We propose meta-reasoning for energy efficiency of multigoal RL, a hardware-aware framework that incorporates low-power preprocessing solutions and meta-reasoning to enable deployment of multigoal RL on tiny autonomous devices. For this aim, a meta-level is proposed to allocate resources efficiently in real time by switching between models with different complexities. Moreover, squeezed-edge you only look once (YOLO) is proposed for energy-efficient object detection in the preprocessing phase. For the experimental results, the proposed squeezed-edge YOLO was deployed on board a tiny drone named Crazyflie with a GAP8 processor that includes eight parallel RISC-V cluster cores. We compared latency and power consumption of squeezed-edge YOLO and a lighter convolutional neural network (CNN)-based model while deploying them separately on board on GAP8. The experimental results show squeezed-edge YOLO is 8× smaller than previous work and consumes 541 mW on GAP8 with inference latency of 130 ms.