GSoC 2026: Question on AI-Assisted Log Diagnosis project (data & scope)

Google Summer of Code
1

Hi everyone,

I’m a prospective GSoC 2026 applicant interested in the
“AI-Assisted Log Diagnosis & Root-Cause Detection” project.

I have a few clarification questions to better understand the scope:

  1. What types of labeled data (if any) currently exist for common failure modes or misconfigurations in ArduPilot logs?
  2. Are there existing datasets, issue references, or log repositories that are typically used as a starting point for this work?
  3. Would an initial prototype be expected to focus more on supervised classification, similarity-based retrieval, or rule-assisted ML approaches?

For background, I have experience with Python-based ML pipelines, data preprocessing, and model evaluation, and I’m looking to align my preparation with the project’s expectations.

Any guidance would be greatly appreciated.

Thanks!
Krishna

1 Like
2

None exist.

None exist.

I guess supervised classification and rule-assisted ML approaches.

2 Likes
3

Thanks for the clarification — that makes sense.

I’ll start by exploring ArduPilot logs and existing failure patterns, and prototype a rule-based labeling pipeline that could later support supervised models.

4

Hi! I would start with a vehicle of choice that you like with SITL to generate logs. You would need to find ways to induce common operational failures in SITL that you can find on our forums. Group these into a set of recurring diagnosis steps and hand label the logs accordingly.

You can then train a simple classifier according to the hand labels and features in the log like sections after certain events or actions in flight, param settings, and then use a small language model fine tuned to explain similar cases for a new log. Would be cool if it could somehow get documentation recommendations too.

Consider some of our previous GSoC projects that involved the use of LLMs for other applications, as this may give you some ideas.

5

Hello! Nate,
I have one question about the feature extraction step.

Different vehicles have different sensors. For example Copter logs there are things like barometer altitude, while Sub or Rover may rely on different sensors such as pressure or wheel data.

When building the feature layer, should the system be designed around a core set of signals common across vehicles, or should it branch into vehicle-specific feature sets for Copter, Plane, Rover, etc?