An OpenAI Gym environment based on PyBullet for learning to control the CrazyFlie quadrotor:

• Can be used for Reinforcement Learning (check out the examples!) or Model Predictive Control
• We used this repository for sim-to-real transfer experiments (see publication [1] below)
• The implemented dynamics model is based on the Bitcraze's Crazyflie 2.1 nano-quadrotor

Circle Task	TakeOff

The following tasks are currently available to fly the little drone:

• Hover
• Circle
• Take-off (implemented but not yet working properly: reward function must be tuned!)
• Reach (not yet implemented)

Here are the (few) steps to follow to get our repository ready to run. Clone the repository and install the phoenix-drone-simulation package via pip. Note that everything after a $ is entered on a terminal, while everything after >>> is passed to a Python interpreter. Please, use the following three steps for installation:

$ git clone https://github.com/SvenGronauer/phoenix-drone-simulation
$ cd phoenix-drone-simulation/
$ pip install -e .

This package follows OpenAI's Gym Interface.

Note: if your default python is 2.7, in the following, replace pip with pip3 and python with python3

We tested this package under Ubuntu 20.04 and Mac OS X 11.2 running Python 3.7 and 3.8. Other system might work as well but have not been tested yet. Note that PyBullet supports Windows as platform only experimentally!.

Bullet-Safety-Gym heavily depends on two packages:

• Gym
• PyBullet

After the successful installation of the repository, the Bullet-Safety-Gym environments can be simply instantiated via gym.make. See:

>>> import gymnasium as gym
>>> import phoenix_drone_simulation
>>> env = gym.make('DroneHoverBulletEnv-v0')

The functional interface follows the API of the OpenAI Gym (Brockman et al., 2016) that consists of the three following important functions:

>>> observation, info = env.reset()
>>> random_action = env.action_space.sample()  # usually the action is determined by a policy
>>> next_observation, reward, terminated, truncated, info = env.step(random_action)

A minimal code for visualizing a uniformly random policy in a GUI, can be seen in:

import gymnasium as gym
import time
import phoenix_drone_simulation

env = gym.make('DroneHoverBulletEnv-v0', render_mode="human")

while True:
    done = False
    x, _ = env.reset()
    while not done:
        random_action = env.action_space.sample()
        x, reward, terminated, truncated, info = env.step(random_action)
        done = terminated or truncated
        time.sleep(0.05)

Note that only calling the render function before the reset function triggers visuals.

To train an agent with the PPO algorithm call:

$ python -m phoenix_drone_simulation.train --alg ppo --env DroneHoverBulletEnv-v0

This works with basically every environment that is compatible with the OpenAI Gym interface:

$ python -m phoenix_drone_simulation.train --alg ppo --env Pendulum-v1

After an RL model has been trained and its checkpoint has been saved on your disk, you can visualize the checkpoint:

$ python -m phoenix_drone_simulation.play --ckpt PATH_TO_CKPT

where PATH_TO_CKPT is the path to the checkpoint, e.g. /var/tmp/sven/DroneHoverSimpleEnv-v0/trpo/2021-11-16__16-08-09/seed_51544

See the generate_trajectories.py script which shows how to generate data batches of size N. Use generate_trajectories.py --play to visualize the policy in PyBullet simulator.

Use Reinforcement Learning (RL) to learn the drone holding its position at (0, 0, 1). This canonical example relies on the RL-safety-Algorithms repository which is a very strong framework for parallel RL algorithm training.

Shows a transfer learning approach. We first train a PPO model in the source domain DroneHoverSimpleEnv-v0 and then re-train the model on a more complex target domain DroneHoverBulletEnv-v0. Note that the DroneHoverBulletEnv-v0 environment builds upon an accurate motor modelling of the CrazyFlie drone and includes a motor dead time as well as a motor lag.

• convert.py @ Sven Gronauer

A function used by Sven to extract the policy networks from his trained Actor Critic module and convert the model to a json file format.

1. Using Simulation Optimization to Improve Zero-shot Policy Transfer of Quadrotors

   Sven Gronauer, Matthias Kissel, Luca Sacchetto, Mathias Korte, Klaus Diepold

   https://arxiv.org/abs/2201.01369

Lastly, we want to thank:

• Jacopo Panerati and his team for contributing the Gym-PyBullet-Drones Repo which was the staring point for this repository.

• Artem Molchanov and collaborators for their hints about the CrazyFlie Firmware and the motor dynamics in their paper "Sim-to-(Multi)-Real: Transfer of Low-Level Robust Control Policies to Multiple Quadrotors"

• Jakob Foerster for this Bachelor Thesis and his insights about the CrazyFlie's parameter values

This repository has been develepod at the

Chair of Data Processing
TUM School of Computation, Information and Technology
Technical University of Munich