Robotics & computer vision
Vision-controlled drone
An ongoing DJI Tello project combining laptop-based flight commands, live camera streaming and local computer vision.
Why I started
I wanted to connect camera observations to physical movement.
My contribution
I worked on the camera stream, perception and control pipeline, bringing together computer vision and device commands.
Current state
In development. The diagram explains the pipeline; no real drone footage or validated autonomous-flight result is available here.
The idea
Translate a high-level instruction, such as finding an object or following a person, into perception and physical actions. The DJI Tello provides the camera and flight platform; a laptop supplies the compute and communicates with the drone over Wi-Fi. The goal is to connect a vision model, local language-model reasoning and programmable flight control in one inspectable system.
The implemented pipeline
I built a laptop-based pipeline combining flight commands, live camera streaming, a local computer-vision model and people/object tracking. The perception side produces detections that can be inspected alongside the camera feed. This creates an observable sequence: receive a frame, identify people or objects, maintain tracking state, and make that state available to the control layer.
People and object tracking belong to the perception pipeline. Turning those observations into a dependable flight policy is the ongoing part: a detection alone does not decide when the drone should move, hold position or stop.
From observations to actions
Keeping compute on the laptop allows local vision and planned language-model orchestration to sit alongside device control. Separating the camera/perception path from the command path makes the system easier to inspect and iterate.
The planned language-model layer translates a natural-language goal into structured actions, while the flight-control layer handles device commands. Flight limits, manual override and emergency stops are part of the behavior being developed and tested.
- Implemented: camera streaming, laptop commands and local perception.
- Developing: object-finding, target-following and agentic operation.
- Planned/developing: safety limits, manual override and emergency-stop behavior.
Current limits and next steps
The project is ongoing. I have not established robust obstacle avoidance, depth estimation, sensor fusion, SLAM or autonomous navigation.
The next step is to develop and test constrained perception-to-control behavior and its safety controls before making stronger autonomy claims. AI coding tools have supported development; the portfolio describes the system I integrated and the implementation boundary.