Reinforcement Learning Quadcopter Agents in RTDynamics & AgileRL Arena#

RTDynamics has teamed up with AgileRL, a London-based Reinforcement Learning specialist, to develop a custom gym training environment. Together, they have successfully demonstrated the training of an AI drone agent using AgileRL's Arena platform integrated with RTDynamics simulation models. The trained agent controls a quadcopter to intercept an incoming drone and defend a moving ground vehicle.

Trained agent will control the tank-spawned drone and intercept an incoming attacker drone

The quadcopter’s flight is simulated with RTDynamics' RotorLib CGF helicopter dynamics model, capable of executing full 6-DoF flight dynamics in vectorized gym environments and faster-than-realtime. The agent interfaces with the quadcopter through RotorLib CGF’s virtual pilot — a comprehensive stack of control systems that exposes multiple modes of interception.

RTDynamics developed the custom gym environment, along with python bindings, which is seamlessly integrated into the AgileRL Arena platform where reinforcement learning takes place.

Users can easily adapt the setup by modifying the helicopter configuration, the environment design (including reward modeling), and the Arena RL parameters to train new agents capable of handling new helicopter configurations and new interception scenarios.

"The training environment will be provided by RTDynamics and can then be uploaded to the AgileRL Arena platform by visiting https://arena.agile.com.

Technology Under the Hood#

RotorLib CGF – Physics-based helicopter flight dynamics with a virtual pilot

Interested in AI topics? Join the RTDynamics AI Newsletter here