CINE
- Role
- Sole researcher
- Timeline
- Aug 2025 – present
CINE (Cinematic Intelligent Navigation Engine) enables a drone to film a moving subject autonomously, keeping the shot steady and well framed with nothing but a small computer on board.
Problem
Self-flying drones can plan a good shot or plan fast enough to fly, but not both on a small onboard computer.
The best camera planners, such as CineMPC, recompute the drone's entire flight path at every step. That takes far more computing power than the small processors on a drone can supply, and without a subject to focus on, those planners produce nothing useful. Fast path planners such as EGO-Planner do run in real time, but they only aim to reach the goal quickly without hitting anything. Additionally, they ignore the way footage looks on camera.
- per hour
- $200–$500
- per project
- $1,000–$5,000
The large cost mostly stems from pilot fees. An autonomous system, on the other hand, can be rented like any other piece of camera gear, creating a cheap system for all.
Solution
Adjust the path the drone already has instead of replanning it
CINE builds on EGO-Planner as a framework and bends the path it produces toward a better shot: holding a steady distance to the subject, placing the subject on the rule of thirds, avoiding glare, and slowing down to cut motion blur. The path is a B-spline, a smooth curve shaped by a handful of control points. Each point only moves its corresponding piece of the curve, and so CINE can fix one interval without recomputing the rest. On board, YOLOv11n finds the subject, Depth Anything 3 estimates its distance from a single camera, and a small neural network corrects the drift in the drone's estimate of its own position.
57%
smaller worst-case error in distance to the subject than EGO-Planner alone
- EGO-Planner0.72 m
- CINE0.31 m
23.6%76.4%
of flight time spent within 10 cm of the target filming distance, before and after CINE
- EGO-Planner23.6%
- CINE76.4%
0.25 m0.05 m
average error in the drone's position estimate, before and after the correction network
- Before0.25 m
- After0.05 m
Flight path results average 15 simulated flights. Detection and depth ran live on the drone's hardware.
Highlights
- Accepted to AIAA SciTech 2027 as a sole-authored abstract
- NYSSEF 2026: third place and the U.S. Air Force award
Stack
- Languages
- Python
- C++
- Robotics
- ROS 2 Jazzy
- RViz
- EGO-Planner
- ArduPilot
- MAVLink
- Hardware
- Raspberry Pi 5
- Coral Edge TPU
- Machine learning
- PyTorch
- torchvision
- Ultralytics YOLOv11n
- Depth Anything 3
- TensorFlow Lite
- ONNX Runtime
- Comet ML
- Vision and data
- OpenCV
- NumPy
- SciPy
- pandas
- Matplotlib