Abstract: Bats demonstrate that advanced interaction with the environment is possible through echolocation, performing activities such as moving between roost and foraging grounds, recognizing specific plants, and capturing prey in the air. For this purpose, they use a highly constrained sensing mechanism, namely emitting and receiving sounds through their mouth and two ears. The timing, intensity and spectral variations embedded in the echoes provide detailed information on re ectors in the surroundings. In this work we present a novel sensor which integrates radar sensing with the biological echolocation paradigm to achieve effcient and robust sensing. Our sensor replicates this mechanism by means of a single emitter, dual receiver setup in combination with an advanced signal processing platform. Radar is chosen as the sensing modality, as it undergoes less attenuation by atmospheric and environmental conditions than either sonar or lidar. The aim of the sensor is to provide an effective sensing mechanism for localization, mapping and autonomous navigation in challenging environments such as industrial or agricultural settings. At the same time, we attempt to minimize the size, weight and power consumption of the sensor, as well as its intrinsic productions cost. In this manner it is possible to integrate the sensor in compact or inexpensive systems. In this work, we explore three main use cases for our sensor; simultaneous localization and mapping (SLAM), autonomous navigation and explicit re ector localization. To achieve SLAM using our sensor we build upon existing work in the field of biologicallyinspired localization. At the basis lies a graph-based SLAM solution which resembles the neural processes related to navigation in rats. The system uses camera images of visited locations as fingerprints and performs place recognition by matching these images to each other to generate a topological map of the environment. We adapt this system to use the readings of our biologically-inspired radar sensor and verify its performance in real-world environments. In the context of autonomous navigation, we focus specifically on sense-and-avoid behavior. This is accomplished by implementing a reactive controller based on the subsumption architecture. This architecture combines multiple behaviors which are designed to handle distinct circumstances and which are governed by an arbitrator. These behaviors cooperate to avoid collisions or recover from them, as well as smoothly curve around obstacles and following the center line of a corridor. We test the system by deploying the controller on a mobile robot and instructing it to wander the oor of an office building while avoiding obstacles. Additionally, we use machine learning to train a neural network to perform a similar functionality. This is done through behavioral cloning, were the network is presented with a recorded data set of instructor velocity commands and recorded radar signals. Training the network derives the relation between the sensor input and velocity output and when deployed on a physical robot, it can navigate the environment without colliding as well. Additionally, explicit re ector localization enables generating point clouds of the environment. This is useful to integrate the sensor with additional SLAM and navigation algorithms. To realize this feature, we analyze the principles underlying biological echolocation; echo timing, interaural differences, and spectral cues. These provide information on distance, azimuth and elevation of a re ector, respectively. We then show how these principles can be transferred to the domain of radar to achieve explicit re ector localization. We derive the theoretical localization capabilities of the sensor through theoretical analysis and experiments. To support the development of the aforementioned functionalities of our sensor, we also introduce a simulator for the pulse-echo sensing modality. The purpose of this simulator is to discover and evaluate potential improvement to the sensor through rapid design space exploration. To achieve this, we combine a state-of-the-art game engine with a computational software suite to generate realistic sensor measurements. The simulated sensor can be mounted on several virtual vehicles, and experiments can be run in different virtual environments and conditions. The simulation balances realism and performance, attempting to produce measurements which contain the information relevant to our applications, while staying as close to real-time performance as possible. To validate the performance of the simulator, we replicate the same experiments of our SLAM and autonomous navigation research and compare the results. Here we show that both simulated components show a strong resemblance to their real-world counter-parts.
Biologically-inspired radar sensing for robotic perception, navigation and SLAM
2022-01-01
Theses
Electronic Resource
English
DDC: | 629 |
Biologically inspired behaviour design for autonomous robotic fish
British Library Online Contents | 2006
|NATURE'S NAVIGATORS -- AN OVERVIEW OF BIOLOGICALLY-INSPIRED NAVIGATION
Online Contents | 2008
|Biologically-Inspired Collective Control for an Autonomous Robotic Arm
British Library Online Contents | 2001
|