This repository contains code to train and run the perceptive bipedal locomotion policies presented in the thesis/paper Learning Perceptive Bipedal Locomotion over Irregular Terrain.

We developed a bipedal locomotion policy using Reinforcement Learning. Our policy improves on the state-of-the-art by integrating (noisy) exteroception, and walking over all kinds of difficult terrains. This repo contains the code to reproduce the results.

There are two models to be trained: a noise free, privileged teacher model, and a noisy, student model that learns to imitate the teacher model.

To train an exteroceptive teacher model:

python train_teacher.py <model name> 60000 exteroception noNorm terrain

Training a teacher model will take 12-36 hours for all 60M timesteps. You can adjust the 60000 for shorter or longer runs. Training progress can be monitored in Tensorboard another terminal by running: tensorboard --logdir teacher_log. The model will be saved every 10% of progress, and can be found in /teacher_log.

Configuration of the environment, curriculum, terrain, etc can be set in config.py.

First find a trained exteroceptive teacher model in /teacher_log that you want to train a student model for. Note the name, e.g.: exp1-60000k-exteroc-noNorm-terrain_0. To train a student model on 1e6 teacher model actions, for 100 epochs:

python train_student.py <student model name> exp1-60000k-exteroc-noNorm-terrain_0 1000000 100

Training a student model takes between 1 min and 1 hour. Training of student models can also be monitored in Tensorboard, and trained models can be found in /student_log.

To run the demo model:

python run.py teacher exteroception demo-model 100000 1

Select the Mujoco windows and hit space to start the sim. Additionally, you can control Cassie with your keyboard (wasdqe) by adding the flag -keyboardControl to these commands. This will spawn an extra terminal displaying the velocity commands (vx, vy, wz). Select this terminal and press any of (wasdqe) to control Cassie!

To run your trained models first find their names in /teacher_log or /student_log.

To run a teacher model that was saved at 60e6 timesteps on 100% curriculum:

python run.py teacher exteroception <teacher model name> 60000 1

To run a student model saved after 100 epochs at 80% curriculum:

python run.py student exteroception <student model name> 100 0.8

Requirements:

• Ubuntu (only tested on 20.04)
• Mujoco210
• cassie-mujoco-sim
• Python env

You can download Mujoco 2.1.0 here

Install it by unpacking the .tar.gz and placing the mujoco210 dir in the ~/.mujoco dir. The path should be: ~/.mujoco/mujoco210

Clone repos:

git clone https://github.com/b-vm/bipedal_walker_terrain
cd bipedal_walker_terrain/src
git clone https://github.com/b-vm/cassie_mujoco_sim

Make sim:

cd cassie_mujoco_sim
make build

Set up Conda env:

conda create --name bipedal_walker python=3.8

Activate Conda env:

conda activate bipedal_walker

Install requirements

cd src
pip install -r requirements.txt

Normally this should be enough to run. However training will be very slow due to some inefficiencies in the recurrent PPO implementation in tensor batching for GPU in sb3_contrib. To make it ~8 times faster, install my fork of sb3_contrib instead. Can be found here. Make sure to install the branch sequence_batching. WARNING: this is not an official release, so expect some tinkering to get it to work.

Clone the forked repo:

git clone -b sequence_batching https://github.com/b-vm/stable-baselines3-contrib.git

Install:

cd stable-baselines3-contrib
pip install .

cd back into /src repo and:

python run.py teacher exteroception demo-model 100000 1

You should now see a Mujoco window with Cassie, and a Matplotlib windows with terrain sensing!

To cite this repo, please cite the paper it belongs to:

@misc{vanmarum2023learning,
      title={Learning Perceptive Bipedal Locomotion over Irregular Terrain}, 
      author={Bart van Marum and Matthia Sabatelli and Hamidreza Kasaei},
      year={2023},
      eprint={2304.07236},
      archivePrefix={arXiv},
      primaryClass={cs.RO}
}

Bart van Marum, Matthia Sabatelli, Hamidreza Kasaei

Work done while at RUG