RI: Small: Learning Resilient Autonomous Flight Behaviors by Exploiting Collision-tolerance

点击翻译按钮获取中文摘要

This project investigates the potential of a new class of aerial robots that exploit collision-tolerance to mitigate risks and challenges during autonomous flight. Modern autonomous unmanned aerial vehicles strive to avoid any possible collision with the environment, a goal which is often hard to achieve and even harder to guarantee. Looking at the animal kingdom, one can observe an alternative paradigm. Small flying insects for example often come in collision with the environment but this does not hinder them from continuing with their lives and tasks. They survive a collision with little to no problem. Inspired by this fact, this project aims to understand the interplay between collision-tolerance and autonomy for agile navigation. Redefining what constitutes “safe navigation” for a small aerial robot, a new class of resilient micro flyers will maximize agility in autonomous flight, while keeping collision risks and possible impacts of a collision below certain acceptable thresholds. This new type of resilient aerial robots will in turn be valuable in a set of real-world applications such as the inspection of hard to access, narrow and visually-degraded environments. This includes but is not limited to the exploration of underground facilities, accessing cargo tanks through manholes and more. In addition, this project strives to contribute into advanced university education and K12 outreach. The latter crucial goals are achieved by close connections between the envisioned research and university classes, and through synergies with established outreach mechanisms to teachers and K12 students in the State of Nevada. To meet these goals, this project builds on top of four research directions. First, it examines a set of alternative collision-tolerant designs for aerial robots by investigating the different advantages and disadvantages of rigid and compliant designs. Second, it aims to design a new “expert” motion planning strategy that explicitly models the risk of a collision and its effect in autonomous navigation. For the latter, the research team will also model the effect of collisions in the ability of the robot to reliably localize. Third, the project team will examine the potential of reinforcement learning methods in the framework of collision-tolerant autonomous flight. Given the hybrid nature of the dynamic phenomena governing collisions and the effect of collisions in the onboard localization functionality of a flying robot, the project envisions a set of contributions in new approaches for learning to navigate that mitigate localization uncertainty through collision resilience and have minimal computational requirements. Eventually, the project aims to facilitate a new class of resilient flying robotic systems capable of supporting multiple real-life applications such as those of industrial and underground inspection. At the same time it aims to introduce leading-edge research to both undergraduate and graduate education activities, alongside strengthening outreach efforts towards K12 students and their teachers.This project is jointly funded by the Robust Intelligence (RI) and the Established Program to Stimulate Competitive Research (EPSCoR).This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.