.
├── README.md
├── results_models                          # explicitely saved models/results
│   └── navigation
├── results_plots                           # analysis results
│   └── navigation
├── rl-crazyflie.code-workspace
├── rl_crazyflie                            # module
│   ├── __init__.py
│   ├── __pycache__
│   ├── assets                              # assets: URDFs
│   ├── envs                                # custom envs
│   └── utils
├── scripts                                 # scripts to analyse/plot results
│   ├── calc_intertia.py
│   ├── err_mag.ipynb
│   ├── tb_logs.ipynb
│   └── viz_log_traj.ipynb
├── test_bal.py                             # train/test balance-aviary
├── test_mo_bal.py                          # train/test multi-objective balance-aviary
├── test_nav.py                             # train/test navigation-aviary
└── test_nav_err.py                         # train/test navigation-aviary with action feedback

1. Install gym-pybullet-drones fork from here arshad171/gym-pybullet-drones:rl-enhancements. The rl-enhancements branch has some nifty enhancements for experiments and multi-objective RL. It is ideal to install the package in "editable" mode. The installation is broken with latest updates of pip, setuptools and wheel. Install these specific versions (recommended by gym-pybullet-drones):

pip install --upgrade pip==23.0.1
pip install wheel==0.38.4 --upgrade
pip install setuptools==66 --upgrade

2. No need to explicitely install stable-baselines3, it should be pulled as a dep by gym-pybullet-drones.

3. Install sb3-contrib, shimmy for the LSTM version of PPO, pip install sb3-contrib shimmy. Ignore stable-baselines3 version conflict.

1. Install gym-pybullet-drones (above).
2. pip install "mo-gymnasium[all]" to install multi-objective gymnasium. Farama-Foundation/MO-Gymnasium.
3. Install the multi-objective version of stable_baselines from here: LucasAlegre/morl-baselines.
4. Install pylibcdd for MORL-baselines from here cdd. The package requires dev tools, check the link for more infomation.

Run xacro stick.xacro > stick.urdf to generate the URDF file. Copy the URDF to gym-pybullet-drones/gym_pybullet_drones/assets/ so gym can access it.

• id = p.loadURDF("path") returns the id of the object (needed for interaction).
• pos, quat = p.getBasePositionAndOrientation(id, physicsClientId=client) returns the position and orientation of the base/root link. The id param is the id returned when loading the URDF.
• rpy = p.getEulerFromQuaternion(quat) converts quat to rpy.

• physical parameters of stick (mass, length, rad): too much => sinks the drone, too light => gets tossed away

• open-loop vs closed-loop: closed-loop introduces too much of control delay => too few RL steps

• rewards:

  • terminate the episode immediately when the stick falls => else noisy reward.
  • a small penatly for moving drifting away from the initial state (along z) => otherwise the agent would learn to simply brace the fall and balance the stick.

• initializing stick at an angle helps after the agent is trained for 1e6

• algorithmic convergence: stochastic policies (PPO, A2C) were found to converge better than deterministic (TD3).

• Model with integral errors has a smoother path, while the other models have a zig-zag path with the agents taking long strides. But the error model does not stabilize well after reaching the destination.

• The magnitude of the error decreases after a few RL steps.

• The model trained on dist (without error) tends to overfit the training scenario (+ve dist), learns maneuvers with max velocity. This does not work well when the direction of wind is reversed or agent is initialized to different point.

• Train longer for better stabilization.

• Errors vs Time

• Init agent at destination and compare the deviations when subject to wind.

• Smoothness $\int |\tau''(t)|^2 dt$, sum of second derivaties.

• Steady-state errors.

• Table highlights.

• Stick balancing

• Navigation improvements

Paper: ArXiv preprint

• Arshad Javeed