2024-06-03 Autonomy Comp Debrief

 Date

Jun 3, 2024

 Participants

  •  

Announcements

  • WARG debrief summary, see comp debrief for more detailed instructions

    • Task 1

      • 35 secs from take off to point Alpha simulator estimate is (43 secs)

      • 3 laps 7500Mah used

      • Weight 9.5 kg

      • Fully autonomous

    • Task 2

      • LZ C

      • Autonomous approach, manual landing

        • Misinterpreted the CONOPS for waypoints

      • Pad was flickering colours but video was good

 Discussion topics

Team

Notes

Team

Notes

UofT

  • Notes

    • Hex

    • RF tower

    • Image
    • Auto stack

      • ROS is 20Hz on Jetson Nano, Mission Planner is 0.5Hz

      • QGroundControl

      • Auto landing uses Geolocation

      • Non Euclidean is a problem, will be 1.5 times the altitude off

    • ↑ Stacks might be outdated as they switch their software stacks three weeks before comp.

      • As far as I know, they are using mission planner for auto mission.

      • YOLOV8

      • range finder lidar

  • Task 1

    • Horizontal to Alpha then ascended

    • Crashed during 3rd lap

    • Minimal damage - destroyed landing gear and cabin, arms and frame are fine

    • Almost vertical oscillation

    • pretty bad tuning

    • mmexport1716647756372.mp4
  • Task 2

    • Crashed and broke cabin, tuned and tried again

    • Crazy oscillations

    • Failsafe and landed during approach (almost hit power line)

Queen’s

  • Notes

    • VTOL

    • Auto stack (Sounds like the most legit one among all teams)

      • ZED X mini angled down (doesn’t work with Pi)

      • Jetson Orin Nano airside, not streaming images

      • ROS for navigation

      • YOLO deployment on Isaac ROS

  • Task 1

    • Flew in quad mode, removed wings

    • Couldn’t get to Alpha, took off and landed

    • Looked stable

  • Task 2

    • Flew in quad mode, removed wings

    • Spiraled downwards during approach

    • Stabilized 2m above ground then crashed nose first

UBC

  • Notes

    • 11kg VTOL

    • Auto stack

      • Custom web based GCS

        • Python 2 Mission Planner Scripts (MPS) running in Docker container

        • Go backend connects to MPS with web sockets

        • Go backend has REST API endpoints

        • React/Svelte frontend connects to backend

        • Theoretically modularized, each module is dockerized, string it up with docker compose, didn’t actually do it

      • Telemetry is RFD 900s

      • Not running auto landing (I got lied to?)

  • Task 1

    • Stalling, couldn’t transition, quad mode was struggling, horizontal prop was on, but wasn’t moving forward

    • Did not get to Alpha

    • Tried 2 auto missions, switched back to manual

    • Used mission planner for path planning

    • They had multiple points around alpha and beta for the turning.

  • Task 2

    • No wings or tail

    • Cardboard tubes for safety

    • Good approach

    • Hovered at end waiting for window

    • Messed up auto landing waypoints and pilot took over manually to land on A

Carleton

  • Notes

    • VTOL (only one that transitioned)

    • 60m/s top speed

    • Auto stack (removed sensors for weight, did not run auto landing)

      • Jetson tx2

      • ZED depth camera

      • Gimbal camera runs object detection

      • Raspberry pi connects to pixhawk forwards MAVLink

      • Wifi

  • Task 1

    • Ascended to 100m, transitioned to horizontal and flew to alpha in 45s

    • Transitioned back to quad at Alpha

    • Turned and flew to Bravo at 60km/h

    • Tried bank the turn at Bravo and crashed

    • Quad mode looked really stable

  • Task 2

    • No wings or tail, cain was not repaired

    • Had a stick in the front to help the pilot with ranging

    • Good approach

    • Hovered at end waiting for window

    • Landed on C but tipped over

    • Calculated PID by hand

ETS

  • Notes

    • Quad

    • Auto stack

      • Tried to research a GNN, didn’t get it working, theoretically more flexible - path doesn’t have to be waypoints

      • Stereo camera (wasn’t mounted, didn’t get to use)

  • Task 1

    • Compass calibration did not work

    • Went above 400ft and judge asked them to land

  • Task 2

    • Compass calibration did not work

    • Did not take off

Alberta

  • Notes

    • Quad with tilted cabin

    • Auto stack

      • YOLOv5 on Pi4 was really slow

      • Streaming images over Wi-Fi, OpenHD

      • Linux with system D using python program to send MAVLink

  • Task 1

    • Took off, got to Alpha, came back and landed - auto mission

    • Kinda wobbly

  • Task 2

    • Auto landed on LZ B

Guelph

  • Notes

    • Fixed wing plane

    • Radio

    • VTX

    • No airside, no mission planner, completely manual

  • Task 1

    • Took off and landed, did not get to Alpha

  • Task 2

    • Took off and landed a few times

Sherbrooke

  • Notes

    • VTOL

    • Massive Wifi antenna - using RFD 900s for comp

    • Auto stack

      • Dronekit, trying to switch to PyMAVLink

      • Airside checks for loiter mode to detect pathing mission ended

      • Landing pad detection with HSV colour detection, not AI model

      • Orange Pi is theoretically more powerful than Raspberry Pi

      • QGroundControl instead of Mission Planner for redundant MAVLink

      • Switching from ArduPilot to ??? because failsafe is weird for VTOL

  • Task 1

    • Crashed on Friday, removed tail and wings to fly in quad

    • Did not get to Alpha

    • Went outside geofence at Bravo and killed the drone

    • Crashed during turn, broke carbon fibre tube

  • Task 2

    • No wings, flying in quad

    • Ascended to waypoint

    • Going to land somewhere

McGill

  • Notes

    • Tailsitter with bending aerofoil

    • Sic radio telemetry link

    • VTX video transmitter

    • Jetty controller duplex connection

    • Auto stack

      • Distance matrix with 1 camera using midas python library - uses an ML model

      • Jetson nano with Python

      • Gazebo simulator

  • Did not fly tasks

UBCO

  • Notes

    • Helicopter

    • Auto stack

      • Pi5

      • Used to use LTE to stream images to groundside

        • Slow, switching to airside compute

      • LiDAR

        • Spinning LiDAR doesn’t work as well in the sunlight

        • Has one directional rangefinders instead

      • YOLOV8 with centering algorithm

      • QGroundControl

  • Did not fly tasks

Manitoba

  • Notes

    • Hex

  • Did not show up

 Action items

@Dylan Finlay to look into ROS - how to deploy models
Look into ZED X mini - any advantages over just using 2 cameras or camera + lidar
Possibly targeting no telem link for auto - use high latency link for controller (GPS data and RTL)
Shelved for later
Test if current proposed hardware setup works - 1 tilted camera (maybe gimbal) + 2 rangefinders
@Yash Gunturi Eshwara Vidya look into simulation software (also for data generation)

 Decisions

  1. Stick with IMACS as is
  2. Support gimbal control system (dynamically change geolocation params)
  3. Targeting airside compute, get that working first before considering ground side
  4. Still look into streaming, logs instead of images?
  5. Stick to dronekit for now
  6. Stick to ArduPilot
  7. Research classical CV methods - stick to AI model because it’s generalizable
  8. Investigate QGroundControl if people are interested, stick with Mission Planner for now