Navix 0.3.6

Accelerated gridworld navigation with JAX for deep reinforcement learning

NAVIX: minigrid in JAX

Quickstart | Installation | Examples | Cite

What is NAVIX?

NAVIX is minigrid in JAX, >1000x faster with Autograd and XLA support. You can see a superficial performance comparison here.

The library is in active development, and we are working on adding more environments and features. If you want join the development and contribute, please open a discussion and let's have a chat!

Installation

We currently support the OSs supported by JAX. You can find a description here.

You might want to follow the same guide to install jax for your faviourite accelerator (e.g. CPU, GPU, or TPU ).

• Stable

Then, install the stable version of navix and its dependencies with:

pip install navix

• Nightly

Or, if you prefer to install the latest version from source:

pip install git+https://github.com/epignatelli/navix

Examples

XLA compilation

One straightforward use case is to accelerate the computation of the environment with XLA compilation. For example, here we vectorise the environment to run multiple environments in parallel, and compile the full training run.

You can find a partial performance comparison with minigrid in the docs.

import jax
import navix as nx


def run(seed)
  env = nx.environments.Room(16, 16, 8)
  key = jax.random.PRNGKey(seed)
  timestep = env.reset(key)
  actions = jax.random.randint(key, (N_TIMESTEPS,), 0, 6)

  def body_fun(timestep, action):
      timestep = env.step(timestep, jnp.asarray(action))
      return timestep, ()

  return jax.lax.scan(body_fun, timestep, jnp.asarray(actions, dtype=jnp.int32))[0]

final_timestep = jax.jit(jax.vmap(run))(jax.numpy.arange(1000))

Backpropagation through the environment

Another use case it to backpropagate through the environment transition function, for example to learn a world model.

TODO(epignatelli): add example.

Cite

If you use navix please consider citing it as:

@misc{pignatelli2023navix,
  author = {Pignatelli, Eduardo},
  title = {Navix: Accelerated gridworld navigation with JAX},
  year = {2023},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/epignatelli/navix}}
  }