Drone Attitude Control in Wind

ISEF Category: Engineering Technology: Statics and Dynamics

Subcategory: Control Theory  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A drone can look stable in a simulator and still wobble hard in real wind. That gap is the whole challenge of sim-to-real transfer. If you can close it, you are working on the same problem that robotics teams face when software leaves the computer and meets the real world.

What Is It?

This project asks a simple question with a hard answer. Can a controller trained in a simulator still fly well on a real quadrotor? A controller is the set of rules that tells the drone how to tilt, spin, and recover when it gets pushed off course.

MuJoCo is a physics simulator. It lets you test flight ideas without risking a real drone. Domain randomization means you train the controller on many slightly different versions of the simulator, so it learns to handle uncertainty, like changes in mass, motor strength, or wind. Think of it like practicing basketball in many gyms with different floors and rims, so you are less surprised in a new court.

PX4 cascaded PID is a standard baseline. PID stands for proportional, integral, and derivative, three terms that help a system react to error, remove bias, and damp overshoot. Your project compares that classic setup with a learned controller and checks which one stays steadier when the air gets messy.

Why This Is a Good Topic

This is a strong science fair topic because you can test it with real numbers. You can measure tracking error, overshoot, recovery time, and stability under wind instead of just saying one drone looked smoother. The project connects to delivery drones, search and rescue, and autonomous flight, where controllers must survive uncertain conditions. You can also learn simulation design, control theory, experimental comparison, and statistical analysis in one project.

Research Questions

• How does domain randomization in MuJoCo affect the real-world attitude tracking error of a quadrotor?
• What is the effect of wind strength on the stability margin of a learned drone controller compared with PX4 cascaded PID?
• Does a controller trained with randomized mass and inertia generalize better to battery-level changes than a controller trained with fixed parameters?
• To what extent does increasing simulator noise improve recovery after sudden orientation disturbances in outdoor flight?
• Which controller, learned or PID, shows lower overshoot and faster settling time during repeated roll and pitch commands?
• How does training on a wider range of motor response models change the gap between simulation performance and flight performance?

Basic Materials

• Sub-250 g hobby quadrotor with programmable flight controller capability.
• RC transmitter and receiver compatible with the drone.
• Laptop or desktop computer for simulation and data analysis.
• MuJoCo installed on a computer that can run physics simulations.
• PX4 firmware and setup tools for the flight controller.
• Wind meter or access to a weather station for recording test conditions.
• Marker cones or tape for a simple outdoor test area.
• Protective goggles and spare propellers.
• Digital logbook or spreadsheet for recording flight results.

Advanced Materials

• Flight controller board that supports custom control interfaces and data logging.
• External motion tracking system, such as Vicon or an optical tracker, for high-precision attitude measurement.
• IMU log access or a companion computer for raw telemetry.
• Calibrated thrust stand for motor characterization.
• Bench power supply for repeatable ground testing.
• Environmental sensor package for wind, temperature, and pressure.
• Safety net or enclosed test space for controlled flight trials.
• University-approved battery testing equipment for consistent power checks.

Software & Tools

• MuJoCo: Simulates drone dynamics so you can train and test controllers before flight.
• PX4: Provides a standard autopilot baseline for comparison against your learned controller.
• Python: Helps you process telemetry, plot error curves, and run statistics.
• pandas: Organizes flight logs into tables that are easy to clean and compare.
• ImageJ: Measures motion from video if you use camera-based tracking instead of onboard logs.

Experiment Steps

1. Define the exact attitude metric you will optimize, such as roll, pitch, or combined orientation error.
2. Choose one baseline controller and one learned controller so your comparison stays clear.
3. Plan the simulation randomization ranges for mass, inertia, motor response, sensor noise, and wind.
4. Design a flight test matrix that varies one real-world condition at a time, such as calm air versus windy air.
5. Decide how you will log and align simulation data, flight data, and video or telemetry data.
6. Select the statistics that will tell you whether the learned controller transfers better than the baseline.

Common Pitfalls

• Training only in one simulator setting, which makes the controller fail when the real drone mass or drag changes.
• Comparing the learned controller and PX4 on different flight paths, which hides whether one controller is actually better.
• Ignoring wind direction, which can make the same test look easy in one run and impossible in the next.
• Using inconsistent battery charge levels, which changes thrust and masks control performance.
• Relying on pretty flight video instead of logged attitude error, which makes the results hard to defend.

What Makes This Competitive

A strong version of this project goes beyond a simple side-by-side flight demo. You would test several randomization strategies, not just one, and show which simulator choices matter most for transfer. You would also use clean metrics, like integrated attitude error, settling time, and failure rate across repeated wind conditions. If you add a careful statistical comparison and explain why the learned controller succeeds or fails, the project looks much more like real control research.

Project Variations

• Compare learned and PID controllers on different drone sizes, such as a tiny whoop versus a sub-250 g quadrotor.
• Test whether training with gust-like disturbances improves recovery better than training with only parameter randomization.
• Analyze sim-to-real transfer using video-based pose tracking instead of onboard telemetry to see how measurement method changes the conclusion.

Learn More

• PX4 User Guide: Read the open documentation for standard drone control and logging features at the PX4 website.
• MuJoCo Documentation: Learn how the simulator models physics and contact at the official MuJoCo docs site.
• MIT OpenCourseWare, Underactuated Robotics: Study control ideas, dynamics, and stability from free course materials at MIT OpenCourseWare.
• NASA Technical Reports Server: Search for papers on autonomous flight, quadrotor control, and disturbance rejection.
• PubMed: Search for review articles on human-safe drone design, flight dynamics, and control interfaces if you want broader context.
• IEEE Xplore: Find peer-reviewed papers on sim-to-real transfer, domain randomization, and quadrotor attitude control through your school or public library access.