Oscillatory Neural Network for Edge Computing: A Mobile Robot Obstacle Avoidance Application
Abstract
Computing sensory data at the edge requires rapid processing with low power consumption. However, treating a large amount of sensory data is not trivial and often requires Artificial Intelligence (AI) algorithms. Conventional AI algorithms are not compatible with edge computation as they demand considerable resources and energy. Also, an additional analog to digital conditioning step is necessary to pre-process sensory data before computation. Thus, ongoing efforts are developed to find solutions offering low-power analog computation. In this work, we develop an oscillatory neural network (ONN) System-on-Chip (SoC) to read, convert and process sensory data. We use the proposed system to perform obstacle avoidance on a mobile robot equipped with eight proximity sensors. To do so, we implement two cascaded ONNs configured as auto-associative memory on an FPGA. We integrate the SoC on the mobile robot running at 0.3m/s and report real-time performances. We report on the reading sensor measurements up to 18ms while the two cascaded ONNs compute only in 40μs, respecting real-time specifications.