Airside compute vs groundside compute

Overview

There has been a massive amount of discussion on airside compute vs groundside compute (multiple years since 2021). This is for the airside system only.

Here is a summary of what has previously been attempted:

Competition

Planned

Outcome

Competition

Planned

Outcome

2021

Groundside compute:

  • Camera: GoPro

  • Transmission: COTS analogue video (not sure what exactly)

  • Compute: Personal gaming laptop

Not run.

Transmission system stopped working, borrowed someone’s Lightbridge 2 but it was not compatible with electrical system.

Fixed wing drone crashed due to balance issue.

2022

Groundside compute:

  • Camera: PiCam

  • Transmission: OpenHD on Raspberry Pi

  • Compute: Jetson

Not run.

Could not get OpenHD working in time. Could not get Jetson working in time. Borrowed the Lightbridge 2 again and used it with the GoPro to observe transmission quality.

2023

Airside compute:

  • Camera: $200 CV camera

  • Transmission: N/A

  • Compute: Jetson

Not run.

Camera arrived very late. Camera and Jetson removed from drone for weight reason. Replaced with groundside compute: Pilot FPV camera, existing transmission system, personal gaming laptop.

Could not get good model quality. Drone did not fly due to VTOL motor overheating issue.

2024

Airside compute:

  • Camera: $200 CV camera

  • Transmission: N/A

  • Compute: Jetson

Not run.

Airside system low priority. Airside system not working in time.

1st place!

2025

Airside compute:

  • Camera: $200 CV camera?

  • Transmission: N/A

  • Compute: Raspberry Pi 5

1st place of course :)

Additional resources

Decision matrix

Criteria

Airside compute

Groundside compute

Criteria

Airside compute

Groundside compute

Control loop latency and reliability

Latency: Small.

Reliability: Very.

Latency: Large, with large variance.

Reliability: The transmission system must be highly reliable, which requires a large amount of development effort.

  • Difficult to test, as there are many possible reasons for failure

The issue is that the compute’s input is images, as frequent as possible and in as high resolution as possible.

Computational power

Limited by weight and power.

Unlimited. The bottleneck is the control loop latency.

Physical effect on drone

Weight: Airside compute adds to this.

Power: Airside compute draws power.

Integration: Must be integrated with the drone’s mechanical and electrical system.

Weight: Transmission system has weight.

Power: Transmission system draws power.

Integration: Must be integrated with the drone’s mechanical and electrical system.

Live health monitoring and recovery

Monitoring: Short range only. Must connect to the airside compute.

Recovery: The airside system must be highly reliable. Any uncaught failure ends the ability for the airside system to operate until the airside compute is rebooted.

Monitoring: All the time.

Recovery: The airside system can be manually restarted at any time.

Ease of failure analysis

Logging: The logging system must be highly reliable. Any failure that is not logged is lost forever.

Logging: All failures are seen regardless of the state of the logging system.

Software development

Compatibility: Airside system must work on both airside compute and developer computers.

  • The airside compute may limit which software versions can be used

Testing: Testing can be done on any developer computer, but it does not guarantee that it also works on the airside compute.

  • Airside compute is a limited resource

    • HITL mitigates this

Compatibility: Airside system must work on any developer computer.

Testing: Testing can be done on any developer computer, but it does not guarantee that the transmission system also works.

  • Transmission system is a limited resource

    • HITL mitigates this

 

 

 

CUDA is abstracted by Pytorch. Even when the Jetson was used, no development in CUDA was done. This is the only information required by members: CUDA and PyTorch

Â