jaxsim 0.1.dev164

A physics engine in reduced coordinates implemented with JAX.

JAXsim

A scalable physics engine implemented with JAX. With JIT batteries 🔋

⚠ This project is still experimental, APIs could change without notice. ️⚠

⚠ This simulator currently focuses on locomotion applications. Only contacts with ground are supported. ️⚠

Features

• Physics engine in reduced coordinates implemented with JAX in Python.
• Supported JIT compilation of Python code for increased performance.
• Transparent support to execute the simulation on CPUs, GPUs, and TPUs.
• Possibility to run parallel multi-body simulations on hardware accelerators for significantly increased throughput.
• Support of SDF models (and, upon conversion, URDF models).
• Collision detection between bodies and uneven ground surface.
• Continuous soft contacts model with no friction cone approximations.
• Full support of inertial properties of bodies.
• Revolute, prismatic, and fixed joints support.
• Integrators: forward Euler, semi-implicit Euler, Runge-Kutta 4.
• High-level classes to compute multi-body dynamics quantities from simulation state.
• High-level classes supporting both object-oriented and functional programming.
• Optional validation of JAX pytrees to prevent JIT re-compilation.

Planned features:

• Reinforcement Learning module developed in JAX.
• Finalization of differentiable physics through the simulation.

Installation

You can install the project with pypa/pip, preferably in a virtual environment:

pip install jaxsim

Have a look to setup.cfg for a complete list of optional dependencies. You can install all of them by specifying jaxsim[all].

Note: if you need GPU support, please follow the official installation instruction of JAX.

Credits

The physics module of JAXsim is based on the theory of the Rigid Body Dynamics Algorithms book authored by Roy Featherstone. We structured part of our logic following its accompanying code. The physics engine is developed entirely in Python using JAX.

The inspiration of developing JAXsim stems from google/brax. Here below we summarize the differences between the projects:

• JAXsim simulates multibody dynamics in reduced coordinates, while brax uses maximal coordinates.
• The rigid body algorithms used in JAXsim allow to efficiently compute quantities based on the Euler-Poincarè formulation of the equations of motion, necessary for model-based robotics research.
• JAXsim supports SDF (and, indirectly, URDF) models, under the assumption that the model is described with the recent Pose Frame Semantics.
• Contrarily to brax, JAXsim only supports collision detection between bodies and a compliant ground surface.
• While supported thanks to the usage of JAX, differentiating through the simulator has not yet been studied.

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

Citing

@software{ferigo_jaxsim_2022,
  author = {Diego Ferigo and Silvio Traversaro and Daniele Pucci},
  title = {{JAXsim}: A Physics Engine in Reduced Coordinates for Control and Robot Learning},
  url = {http://github.com/ami-iit/jaxsin},
  year = {2022},
}

Maintainers

	@diegoferigo

License

BSD3