Machine Learning Drone
For this project, my team and I built a drone from scratch designed to detect and track people without needing a ground station or external laptop. By moving all processing onboard, we created a fully automatic system with effectively "infinite range." As the lead for the design and hardware integration, I performed the lift and drag calculations to ensure flight efficiency and conducted center of gravity analysis to balance the drone’s weight perfectly. I was responsible for the material selection and the complete design of the chassis, ensuring it could house the electronics securely. On the technical side, I handled the electrical integration, successfully wiring and troubleshooting the power systems to link our AI hardware with the flight controller and motors. My work ensured that the drone was aerodynamically stable, lightweight, and capable of running complex machine learning tasks entirely on its own.
Because we chose a quadcopter design, I used the relationship between propeller disk area and static thrust to determine the drone's scale. I estimated a maximum takeoff weight of 2.5kg by accounting for the necessary components such as the NVIDIA Jetson, propulsion system, battery, etc. Using the formula for static thrust:
Thrust = Area x Pressure Differential
I determined that a propeller diameter of 12 inches was necessary to achieve a 2:1 thrust-to-weight ratio, which is the industry standard for stable flight. This calculation dictated the minimum arm length and overall footprint of the drone. From there, I moved into the CAD phase, designing a chassis that balanced the necessary spacing for electronics with the structural requirements of a 2.5kg aircraft.
The development of the chassis began with a comprehensive requirements analysis to define the mechanical and electrical architecture. I identified all mission-critical components, including the motors, airframe arms, NVIDIA Jetson, ESCs, and the Pixhawk flight controller, to ensure a cohesive system layout.
During the conceptual phase, I produced several design iterations to determine the most efficient use of space and weight. A key priority was optimizing the drone's stability, which I achieved by intentionally lowering the center of gravity through strategic component placement. Once the layout was finalized, I transitioned the project from initial sketches to a high-fidelity 3D CAD model. This allowed for precise spatial verification and part integration before I moved into the prototyping stage, where I conducted physical testing for fitment, tolerances, and structural integrity.
The chassis was fabricated using ABS due to its superior impact resistance and high thermal stability compared to standard filaments. This material choice ensured the frame could withstand both the heat generated by the NVIDIA Jetson and ESCs and the structural stresses of flight. For the propulsion system, I integrated carbon fiber structural tubing to serve as the drone arms, providing a high strength-to-weight ratio. These arms housed the repurposed DJI E600 motors, which were sourced and tested for reliability.
To guarantee a minimum flight endurance of 10 minutes, I conducted a power budget analysis. This involved measuring the current draw from the motors at hover and the peak power consumption of the Jetson Orin during AI inference. By calculating the total milliamp-hour (mAh) requirements, I was able to specify a battery capacity that supported both the heavy compute load and the lift requirements. As a critical safety measure, I included a fusible link into the primary power distribution circuit. This provides an essential fail-safe to protect the expensive onboard electronics from damage in the event of a short circuit or overcurrent condition.
6s 22.2V 6000mAh Lipo battery
A primary challenge in the integration phase was optimizing the internal packaging to house all necessary hardware within a minimal footprint. I designed the internal layout to be as compact as possible to reduce the overall profile while maintaining dedicated airflow pathways. This thermal management strategy was essential to prevent the high-performance electronics from throttling or overheating during extended autonomous missions.
In addition to thermal considerations, I allocated specific volume for clean wire management to prevent interference with the flight controller's sensors. To increase system reliability, I implemented rigorous insulation protocols for all exposed connections and terminals. By utilizing Kapton tape for its high dielectric strength and thermal resistance, I effectively minimized potential failure modes related to electrical shorting or vibration-induced wear. This attention to detail ensured a robust electrical architecture capable of withstanding the rigors of flight.
On the electrical and software side of this project, we created our high-performance machine learning drone by integrating an NVIDIA Jetson Orin with a Pixhawk 2.4.8 flight controller for fully onboard autonomous tracking. On the software front, we optimized a custom YOLOv11 algorithm to achieve a 300% increase in detection speed; this was accomplished by narrowing identification parameters specifically to human targets and engineering a logic gate to skip compute cycles on stagnant frames. To meet a strict $250 budget, I sourced DJI E600 motors and ESCs. My work bridged the gap between computer vision and hardware execution, ensuring that the Jetson’s real-time processing translated into stable, responsive flight commands through a custom-configured electrical architecture.
Videos of drone:
This project serves as a foundational platform that our team continues to evolve through iterative development. Our future objectives are focused on increasing flight time by light weighting our current drone chassis design. We also plan to advance the computer vision system by refining the YOLOv11 implementation to include multi-object tracking and target re-identification. Furthermore, we are working toward multi-axis autonomy to enable 3D path planning and advanced collision avoidance. I remain committed to the continuous improvement of this platform and will update this documentation as we reach new milestones.
Below is a link to a website regarding this project and the team who made it possible.