GSoC 2026 Proposal Interest | AI for Root Cause Detection in ArduPilot Logs

Google Summer of Code
1

AI Assisted Log Diagnosis & Root Cause Detection for ArduPilot

 Harshita Adlakha: https://www.linkedin.com/in/harshita-adlakha-48981235b/

Modern autonomous systems generate massive volumes of telemetry logs but extracting actionable insights from them remains a bottleneck. In complex flight systems like ArduPilot, diagnosing failures often requires expert level manual analysis of logs, parameters, and flight conditions.

For GSoC 2026, I am deeply interested in building an AI assisted diagnostic system that transforms raw flight logs into intelligent, explainable root cause insights.

Problem Statement

Flight logs contain rich information sensor readings, control signals, parameter states but:

  • Debugging is time intensive and expert dependent

  • Failures often arise from non obvious parameter interactions

  • There is no automated system to map log patterns → root causes → fixes

This creates a critical need for a system that can:

Automatically identify failure patterns and suggest corrective actions with confidence.

Proposed Solution

I propose developing a hybrid ML based diagnostic engine that combines:

1. Log Pattern Classification

  • Use supervised learning models (e.g., Random Forest, XGBoost, or lightweight neural networks)

  • Input: Structured features extracted from ArduPilot logs

  • Output: Probable failure categories (e.g., EKF issues, vibration problems, GPS glitches)

2. Retrieval Augmented Diagnosis

  • Build a knowledge base from:

    • Historical logs

    • Known issues

    • Parameter configurations

  • Use similarity search (e.g., embeddings + vector DB) to retrieve:

    • Similar past failures

    • Proven fixes

3. Root Cause + Fix Recommendation Engine

  • Combine classification + retrieval outputs

  • Generate:

    • Root cause hypothesis

    • Suggested parameter changes or actions

    • Confidence score

    • Evidence links from logs

System Architecture (High Level)

 Raw Logs → Preprocessing → Feature Extraction
            ↓
     Classification Model ──┐
                           ├──→ Diagnosis Engine → Output
   Retrieval System  ──────┘
                           ↓
         Root Cause + Fix + Confidence + Evidence

Technical Approach

Data Processing

  • Parse .bin or .log files using ArduPilot tools

  • Extract:

· Time-series features (IMU, GPS, EKF, etc.)

· Event markers and anomalies

  • Normalize and segment logs into meaningful windows

Feature Engineering

  • Statistical features (mean, variance, spikes)

  • Temporal features (rate of change, drift)

  • Domain specific indicators (EKF variance, vibration thresholds)

Modeling Strategy

  • Classification Models:

    • Baseline: Logistic Regression

    • Advanced: XGBoost / LightGBM

  • Retrieval System:

· Sentence transformers for embedding log summaries

· FAISS / vector DB for similarity search

Explainability

  • Feature importance (SHAP or similar)

  • Highlight exact log segments influencing predictions

Expected Output

For a given log, the system will return:

  • Probable Root Cause (e.g., EKF divergence due to GPS glitch)

  • Suggested Fixes (e.g., adjust EKF parameters, check GPS module)

  • Confidence Score (e.g., 87%)

  • Evidence (specific timestamps or log segments)

Tools & Technologies

  • Python (core development)

  • ML libraries: Scikit learn, XGBoost, PyTorch

  • Log parsing: ArduPilot MAVLog tools

  • Vector DB: FAISS

  • Visualization: Matplotlib / Plotly

Alignment with ArduPilot Goals

This project directly contributes to:

  • Reducing debugging time for developers and users

  • Making ArduPilot more accessible to non-experts

  • Enhancing reliability of autonomous flight systems

  • Leveraging AI for real world aerospace challenges

My Preparation & Next Steps

To ensure a strong contribution, I am:

  • Studying ArduPilot log structure and parameters

  • Practicing Git/GitHub workflows through open source contributions

  • Building small prototypes on:

    • Time-series anomaly detection

    • Log classification

  • Exploring past ArduPilot issues and debugging workflows

Why This Project Matters

As autonomous systems scale, intelligent observability becomes critical.

This project is not just about automation it’s about:

Turning raw data into engineering intelligence.

By combining machine learning with domain knowledge, we can move from:

  • Reactive debugging → Proactive diagnosis

  • Manual analysis → Intelligent assistance

I am excited about contributing to ArduPilot through this project and working under the guidance of experienced mentors to build something impactful, scalable, and genuinely useful for the community.

#GSoC2026 ardupilot #OpenSource #MachineLearning #ArtificialIntelligence #AerospaceEngineering #AutonomousSystems uav python #BuildInPublic

1 Like