Information-Based Supervised Learning of In-Proximity Effects for 3D Distance Estimation and Collision Avoidance

Abstract

In-proximity effects (IPE) in 3D, specifically in-ground, in-ceiling, and in-wall effects, experienced by a rotary-wing aerial robot as it flies near obstacles are leveraged for obstacle distance estimation and collision-free motion control. Onboard motor commands and inertial measurement unit (IMU) signals are processed to enable the robot to essentially “feel” the presence of nearby obstacles through aerodynamic interactions. The physics of IPE, along with Shannon information, are used to tailor the input space and train a deep neural network (DNN) to estimate the distance to ground, ceiling, and wall features. Simulation and physical experimental results demonstrate reliable and robust obstacle detection and collision avoidance with a median distance estimation accuracy of 93.35%, 89.22%, and 90.67% for ground, ceiling, and wall, respectively. This new form of “sensing” is useful in environments with fog, smoke, dust, rain, or snow, where traditional proximity sensors and vision-based systems struggle to detect obstacles and determine distance.