RL Aerial-Target Tracking

End-to-end reinforcement-learning policy for tracking aerial targets. Policies trained in PyTorch via Stable Baselines3 against an OpenAI Gym environment wrapped around AirSim and Gazebo, with ROS 2 as the runtime bridge to the tracking platform. Beats classical PID baselines by 5× at 10 m/s target velocity.

What it does

The agent observes the relative state of a moving aerial target and outputs continuous control commands to a tracking platform. The training loop runs through a custom OpenAI Gym environment that drives AirSim (for high-fidelity drone dynamics) and Gazebo (for ground-truth physics), with Stable Baselines3 handling the PPO/SAC training. ROS 2 is the bridge between the trained policy and the runtime tracker.

The headline result, 5× tracking-error reduction over tuned PID at 10 m/s target velocity, is the operating point where PID controllers start to break down because the linear gain assumptions stop holding. The RL policy learns the nonlinear correction implicitly.

Stack

PyTorch (policy backbone), Stable Baselines3 (RL training algorithms), OpenAI Gym (environment interface), AirSim (high-fidelity drone sim), Gazebo (physics and ground-truth state), ROS 2 (real-time messaging between policy and platform), Python.

Where it fits

This work was done in the ADAMS Lab alongside the IROS 2026 submission. Where the swarm work focuses on multi-agent coordination without a central coordinator, this one focuses on the single-agent control problem: high-rate, nonlinear, hard for hand-tuned controllers.

Status

Shipped. No public repository or paper yet. Happy to walk through the setup or share artifacts on request.