parkour-training

This is one of two main repositories for the parkour project at Roboy. Our initial goal was to provide an environment and initial solution for the parkour challenge.

After some time and direction changes, we conctentrated on trying to pass through the bouncy-catle track untill the ramp and after using a simple policy, to experiment with training a motion-capture mimicking solution.

Repositories and code stucture

Our environment, initial training experiments and simple walking demo can be found here.

For the advanced experiments with learning motion-capture based and compositional policies, refer to parkour_learning

For the simple deployable environment in pybullet-gym, refer to parkour-training-starting-kit

Dependencies and installation

We use OpenAI Gym, PyBullet, PyBullet Gymperium and BAIR's deep learning library rlpyt for the training and the environment.

Installing pybullet

pip install pybullet

Installing OpenAI Gym

git clone https://github.com/openai/gym.git
cd gym
pip install -e .

Installing PyBullet Gymperium

git clone https://github.com/benelot/pybullet-gym.git
cd pybullet-gym
pip install -e .

Installing rlpyt

git clone https://github.com/astooke/rlpyt.git
cd rlpyt
pip install -e .

Installing gym_parkour

And finally, to install the content of this repository you need:

git clone https://github.com/Roboy/parkour-training
cd parkour-training
pip install -e .

Repository structure

This repository mainly consist of our primary environment ParkourChallenge-v0, which is implemented, following a hierarchical structure with easily swappable robots and scenes.

All our robots are based on the ParkourRobot abstract class that you can find in gym_parkour/envs/parkour_robot.py, which demands the implementation of state calculation, actions, updates, reward calculation, etc.

The canonical robot used for the demo is a basic pybullet humanoid with all the methods and specifications implemented in the gym_parkour/envs/humanoid.py.

The main environment is implemented in gym_parkour/envs/parkour_gym.py and follows the BaseBulletEnv specifications from pybullet-gym.

Examples

For a pre-trained model that used the proximity to the goal as the main reward, we've created a demo, that changes the goal positions to follow the track progression.

You can watch pretrained humanoid running through the parkour track like this:

cd parkour-training
python3 enjoy_humanoid

Bugs

pbullet migth throw an error if several environments run in parallel: pybullet.error: Not connected to physics server.

for fix see here: https://pybullet.org/Bullet/phpBB3/viewtopic.php?t=12722