All work

University of Zürich · Robotics and Perception Group · Sept. 2025

Deep-RL Vision-Based Drone Control

A PPO controller for vision-based target following under partial observability, transferred from simulation to a real drone with stable altitude and stand-off control.

From pixels to flight commands

This project asks a drone to follow a target using visual observations while keeping a safe distance and stable altitude. The policy must act under partial observability and tolerate the visual and dynamic mismatch between a simulator and a physical aircraft.

The control stack

I trained a Proximal Policy Optimization (PPO) controller around a double-sphere bounding-box simulator, field-of-view-aware rewards, smooth command generation, and domain randomization. The perception-to-control interface was kept intentionally compact so the same policy could move from simulation into the ROS2 flight stack.

Real-world result

The deployed controller followed the target safely while holding altitude with a mean error of −0.10 m. The demo and report cover the reward design, sim-to-real choices, flight architecture, and failure cases.