McGill Robotics Drone | Flight Readiness Review | AUVSI SUAS 2018

my name is Molly I'm a third year honors in your knowledge ecn Abigail I'm the project manager of the dream project this is my third year with McGee in robotics and my first year was a joint project on the flightline I will be in charge of coordinating between the flight line members and the judges as well as taking footage hi my name is Anna and I am a graduating electrical engineering student I joined Vigo robotics five years ago and have been leading drones software efforts for the past three years this year I architected and oversaw all software development on the flight line I'll be supervising all ground station operations and making sure everything goes smooth hi my name is babe and I'm graduating telecom engineer student I joined robotics five years ago and I have in the drone – green the past two years I have design and implement the aircraft the electrical system ensuring safe under level operations I will be at the flatland to debug potential issues and to make emergency repairs hi my name is shot and I'm a second year mechanical engineering student and miquelon this is my second year on the team and I am the mechanical division lead of the front project I spent the past year working on designing and building our competition effort on the flight line I've been in charge of overseeing the UAV setup and disassembly hi my name is Jessa I'm a third year electrical engineering student I've been working with negative robotics for three years and this is my second year on the drone project this year I went primarily on the software and more specifically the obstacle avoidance and that's exactly what I'm gonna be monitoring on the flight life hi I'm Esther yashi I'm a graduating electrical engineering student I'm on the software section of the drone team working on computer vision hi my name is John Waite and I'm a second year student of Electric in general I have a year of experience if you Ravens on the drone software sub team I worked on the internal client and the mission planner and those are the things I'm Matthew Phillips I'm in my third year of Chemical Engineering and a minor in software it's my first year on the team and I have experience in aerodynamics and carbon fiber manufacturing I'm on the mechanical sub team I've helped with various things including helping with the design of skis for winter testing on the flightline I'll be helping with plane setup I am this is my first year on the team I've had no previous experience with UAVs I'm in the mechanical sub team and this year I'm a daily basis University in mechanical engineering I am part of the mechanical team of drone and I've been working on fabricating designing and manufacturing some parts I have some experience in aerial frames that have required on my own but this is my person I'm very excited wheelie 4 is a fixed-wing twin boom pusher aircraft and mega robotics entry for the competition it is a ready-made RC anaconda that has been modified according to our needs and fitted with all the required hardware including an on-board computer equipment camera and all of the necessary avionics components are mounted on detachable 3d printing base allowing for modularity and diese faxes 24 employs a 6-cell 13 amp hour lithium polymer battery to power a 365 kV motor with a 15 by 10 inch propeller an Intel nuke is used to as the companion computer alongside of pixhawk 2.1 autopilot the software system runs on the PX or flight stack and the robot operating system the system was designed to operate fully autonomously and does not rely on any communication to the ground to accomplish competition tasks everything runs on board the aircraft and the ground operators will only serve to guide or modify its behavior when they see fit Guney four will attempt to perform all competition tasks our degree of confidence was each task shown in this table is based on the extent of testing done and performance during tests our development process also relied on a software in the loop simulator integrated with gazebo it is capable of simulating all mission tasks and can randomize them to all to our liking moreover a continuous integration server automatically runs all units and simulated mission tests before any change can get committed as such we could test and iterate very quickly and minimize flight test crashes due to easily avoidable problems in addition the team relied on a smaller surrogate airframe goony 3 for testing software without risking the competition airframe over the course of the year it accumulated multiple hours of flight and proved instrumental in validating newly written code finally only once all of our automated tests had passed and once we were confident with its performance on good III were we ready to conduct similar tests on our competition plane only four has consistently flown every weekend over the course of the year winter testing was principally done using unique three was include creased guna for testing later in the year all of our flights were fully autonomous except when testing new airframe setups additionally since we use an off board planner the only required auto pilot side tuning was letting the controller flight angles everything else is adaptive this year we managed to get a very good obstacle avoidance scheme that can avoid both moving and stationary obstacles in real time this was achieved with a local planner known as the dynamic when to approach or DWA this planner works by simulating set of controls at every time step then choosing the best one based on criteria that we provide we decided to penalize any control that went too close to a moving or stationary obstacle on top of this we were ordered controls that tended to face the goal the end result of this is that the plane will always go towards a goal but will always automatically dodge any obstacles in its path to further improve our results we predict moving obstacles trajectories nearly perfectly ten seconds into the future and incorporate the predictions into DW simulations we have extensively tested our planner on a multitude of randomly generated moving and stationary obstacles with 100% accuracy we have also tested this on our goon III test life albeit not as exhaustively another advantage of our DW planner is its increased waypoint accuracy performance as we no longer need to rely on the px4 is LM controller this gives us complete control over how accurately we want to hit our waypoints for now we set our acceptance radius to 15 feet which our DW finder will continuously attempt to it as such we can hit 100 percent of waypoints within 15 feet while avoiding obstacles we have tested this exhaustively both in simulation and in air on the meter using the point gravely three camera with an 8 millimeter lens and flying at an altitude of 120 feet the smallest possible target 1 foot wide would take up 60 pixels our tests have deemed this sufficient moreover to ensure the best image quality we have tuned our gimbal controller to stabilize the footage as much as possible we trained a sliding window classifier which extracts features from a window using a convolutional neural network and classifies whether that window contains part of a target the windows with a positive prediction are used to compute a box around the object we trained and evaluated the detection network on generated data obtained using the gazebo simulator our validation accuracy on generated data is around 98% we anticipate that our models performance will be much lower on real-world death which is why we have implemented a domain adversarial model a few real targets were crafted in order to collect we however due to difficult weather conditions not enough images with the point gray camera were collected yet evaluation of target attributes is achieved through a combination of k-means clustering open CVS contour finding functions and tesseract OCR which have similarly promising results on our simulated data over the next week we will be collecting more real footage for our domain adaptation based on our aircrafts estimators average converged covariance our estimated standard deviation would be approximately 5 feet this means that we have over a 99% chance of localizing within 15 feet of the target our gazebo simulations on thousands of generated targets concise with this measurement error but we have yet to validate this in real the air delivery system is mounted on the bottom of the plane and drops the bottle in a circle that's service training lunch juice the bottle breaks on impact without needing any modification we have a simulation going for the air drop at all times which estimates the location the bottle will hit when the estimated position of the air drop is close enough to the target we will release the bottle we assumed constant wind speed and direction during the fall for the estimation we were able to easily test this in our single address shown here cannot test ruggles we were able to deliver the air drop package within 15 feet of the target to 80% so far our full mission tests have been limited to simulations since goony 4 cannot take off or land in Canadian winters moreover the extreme cold limits goony threes battery life to 3 minutes which is insufficient for a complete mission instead a number of isolated tests were carried out to validate the design this allowed us to gather valuable operational knowledge as well as identify any flaws with the initial setup problems were identified with the front landing gear and camera game consequently the landing gear was redesigned and replaced and the gimbal was remodeled to have better balance now that we have validated boonie fours mechanical design and that weather finally permits we will be focusing on more exhausted full mission tests until competition based on our full mission tests we estimate that we will get 40% for the mission time criteria as we have averaged 25 minutes while performing the full range of mission tasks we also do not expect any penalties or timeouts as we never go over the full 45 minutes competition period for the autonomous flight task our tests show that we should receive full marks for both the autonomous flight and the Waypoint capture sections as we never go out of bounds while hitting 85% for waypoint accuracy moreover our tests show that we should expect full marks for the obstacle avoidance sections of the marking scheme along the same line we expect 75% or more for all object detection classification and localization tasks as for air delivery our tests have shown an average of 80% accuracy finally due to us not suffering any major crashes and due to last year's performance we expect 80% for operational excellence as well overall we expect our total emission score to be over 80%

This is McGill Robotics Drone project’s flight readiness review for the AUVSI SUAS 2018 competition.

Music: Drops of H20 by J. Lang

Leave a Reply

Your email address will not be published. Required fields are marked *