Python interface for Dojo

This package is a Python wrapper for the differentiable simulator Dojo.

• arXiv preprint: https://arxiv.org/abs/2203.00806
• site: https://sites.google.com/view/dojo-sim
• video presentation: https://youtu.be/TRtOESXJxJQ

Included are interfaces to PyTorch and JAX.

Quickstart

This example simulates a pendulum for 1 time step.

import dojopy 
from julia import Base
from julia import Dojo as dojo

# get an environment
env = dojo.get_environment('pendulum')
dojo.initialize_pendulum_b(env.mechanism, angle=0.0, angular_velocity=0.0)

# get state
x1 = dojo.get_minimal_state(env.mechanism)

# random control
u1 = Base.rand(nu)

# simulate one time step
dojo.step(env, x1, u1)

Installation

Using Dojo with Python requires a number of installations in addition to dojopy. Below are two options for installing all dependencies.

Using Docker (recommended)

1. Clone this repository:

git clone https://github.com/dojo-sim/dojopy

2. Install Docker (https://docs.docker.com/engine/install/)

3. Enter the repository folder, build the Dockerfile and tag the image as dojopy: (this step may take a few minutes)

cd dojopy
docker build --tag dojopy .

4. Open a bash shell in your Docker container

docker run -it -v /absolute/path/to/dojopy:/dojopip dojopy bash

(-v /absolute/path/to/dojopy:/dojopip allows you to synchronize files from your folder /absolute/path/to/dojopy to your docker image)

5. You can now run dojopy inside your Docker image! Inside the shell of your Docker image opened in step 4., run

python3 quick_start.py

Manual (advanced)

Calling Dojo from Python requires:

• dojopy: this wrapper
• Julia v1.6+
• Dojo.jl: the actual simulator
• PyCall: interface between Julia and Python
• custom Python binary: this is required to make calls to Dojo fast and efficient

Below we walk through each of the required installation steps:

Get dojopy

1. Clone this repository:

git clone https://github.com/dojo-sim/dojopy

(for now, soon via pip)

Custom Python installation To make calls from Python to Dojo efficient requires a custom Python installation.

2. Install pyenv

   • Installation guide for Ubuntu
   • Installation guide for MacOS
   • Installation guide for Windows

3. Use pyenv to build your own Python

   In ~/.pyenv run:

   PYTHON_CONFIGURE_OPTS="--enable-shared" pyenv install 3.6.6

   to create a custom binary.

   • We call this python binary the custom_python. It's located at path/to/custom_python e.g., /home/user/.pyenv/versions/3.6.6/bin/python3
   • This step is needed because PyJulia cannot be initialized properly out-of-the-box when Python executables are statically linked to libpython. This is the case if you use Python installed with Debian-based Linux distribution such as Ubuntu or installed Python via conda. More details about this here.

4. (Optional, Recommended) Create a virtual environment linked to custom_python

   • In your shell run:

   path/to/custom_python -m venv /path/to/new/virtual/environment/my_env

Julia installation

5. Install the Julia programming language (v1.6+ recommended) [Julia Download page]

6. Install PyCall

   • Specify the Python version to be the custom_python.
     • e.g. ENV["PYTHON"] = "/home/user/.pyenv/versions/3.6.6/bin/python3"
     • Pkg.build("PyCall")

7. Open the Julia REPL and install the Julia package Dojo.jl: (type ]):

pkg> add Dojo

Python setup

8. In your virtual environment, install: pyjulia, the interface that lets you call Julia code from Python.

   • Activate your virtual environement, then run:

   python3 -m pip install julia

9. In Python run:

import julia
julia.install()

to finish the pyjulia setup.

We can now call Dojo from Python!

Documentation

See the Documentation for using Dojo.

Performance

When Dojo is called from a python script, e.g. python3 ... Julia will just-in-time compile the solver code which will slow down the overall execution. For larger problems it is advisable to solve a mini problem first to trigger the JIT-compilation and get full performance on the subsequent solve of the actual problem .

License

This project is licensed under the MIT License - see the LICENSE.md file for details.

Citing

@article{howelllecleach2022,
	title={Dojo: A Differentiable Simulator for Robotics},
	author={Taylor, A. Howell and Le Cleac'h, Simon and Kolter, Zico and Schwager, Mac and Manchester, Zachary},
	journal={arXiv preprint arXiv:2203.00806},
	url={https://arxiv.org/abs/2203.00806},
	year={2022}
}

How To Contribute

Please submit a pull request, open an issue, or reach out to: thowell@stanford.edu (Taylor) or simonlc@stanford.edu (Simon)