Goal

Additional support and instructions

• Current bootcamp is very open ended and is difficult to get started (especially if new to Python and/or neural networks

More complexity/creativity

• Current bootcamp is very straightforward: Optimize on dataset

Less computationally intensive

• Current bootcamp requires a lot of time for training

Proposals

Landing Pad Search

Bootcamper task: Find all landing pads in the simulated area, using the simulated drone and the given landing pad model.

Landing pad model: Same model used for the 2022 competition.

• Ultralytics for inference

Simulator:

• Premise: The simulated drone is a quadcopter with a downward facing square camera at a constant height of 60m and max speed of 5m/s .

• Model:

  • Provided: Given the drone’s position, map tile images are stitched together and then cropped. The resulting image and position is put into the output queue. This simulates the camera.

  • Bootcamper task: Get the images from the input queue and use them as input to the model, which may find landing pads and output bounding boxes. Call the Geolocation code with the bounding boxes and drone position.

  • Provided: Bounding box boundaries/centres are converted from image space to world space using drone position. This is a simple scaling and translation operation.

  • Bootcamper task: Use some logic and control the drone by putting a command into the output queue.

• View: Created from the model. Minimap with drone position and current command, current view of camera with bounding box outlines and coordinates.

• Controller:

  • Commands are asynchronous messages to the drone controller. Available commands:

    • Set destination position.

    • Halt at current position.

    • Land at current position (requires drone to be halted, ends the simulation).

  • The drone uses the latest command (i.e. a new command will supersede the current command).

Write-up:

• Documentation on how the decision logic is made.

• Optional: How to handle the following:

  • Detecting is not 100% accurate.

  • Locating landing pad is not 100% accurate.

Alternatives:

• Restrict drone within area (fail if drone moves outside of area).

• Landing pads inside restricted area (fail if drone attempts to land there).

• Won’t do: Geolocation is an additional task that needs to be calculated by hand by bootcamper.