RealBrandonChen / DroneSim

Reinforcement learning for an AirSim quadrotor implemented in Unity

DroneSim

Reinforcement learning for path following of an AirSim quadrotor implemented in Unity city environment

Drone in training:

Figure1. training started	Figure2. 4,000 time steps	Figure3. 12,000 time steps

Getting Started

• Download the Unity package containing the customized city environment and the AirSim drone;

• Add the python files in Path_Following to the AirSim/PythonClient/reinforcement_learning you've cloned;

• You can

  • Directly load the pre-trainined model by hitting python Model_Load.py in the terminal, and you'll see the drone following the city road

  • Train your own model by python dqn_drone.py, your trained model is saved as "best_model.zip"

Implementation Explanation

Code snippet credit to AirSim/PythonClient/Reinforcement_learning/drone_env, and the reward function is as following:

def _compute_reward(self):
    thresh_dist = 7
    beta = 1
    x = -240
    y = 10
    z = 200
    pts = [
        np.array([x, y, z]),
        np.array([-350, y, z]),
        np.array([-350, y, 150]),
        np.array([-350, y, z-100]),
        np.array([-350, y, z-200]),
    ]
    ...
    ...
    if self.state["collision"]:
        reward = -100
    else:
        dist = 10000000
        for i in range(0, len(pts) - 1):
            dist = min(
                dist,
                np.linalg.norm(np.cross((quad_pt - pts[i]), (quad_pt - pts[i + 1])))
                / np.linalg.norm(pts[i] - pts[i + 1]),
            )

        if dist > thresh_dist:
            reward = -10
    ...
    ...
    done = 0
    if reward <= -10:
        done = 1

    return reward, done

The tuple of the coordinates represents the central line of the city road. The dist in the reward function computes the twice distance between the realtime drone and the central line comprised by the points. The distance computing is reperesented as following picture:

Future Work

• Generate images data with imitation learning

  A trained policy by cross-modal representations has been achieved by Rogerio Bonatti. The imitation learning data is generated for passing through the drone racing obstacles. The path following task should also work applied with the generated imitation learning data.

• Implement the trained model in the real drone

  Transfer the simulation algorithm to real-world platform.