| <!-- | Grid | Diff Drive | Continous | | ---- | ---- | ---------- | --------- |>

JAXMAN is a JAX-based library for multi-agent navigation. Our library can create environments with three different dynamics.

$ python -m venv .**venv**
$ source .venv/bin/activate
(.venv) $ pip install -e .[dev]

$ docker-compose build
$ docker-compose up -d dev
$ docker-compose exec dev bash

$ docker-compose up -d dev-gpu
$ docker-compose exec dev-gpu bash

and update JAX modules in the container...

# pip install "jax[cuda]" -f https://storage.googleapis.com/jax-releases/jax_releases.html

• 1. Environment
• 2. RL Agent

Test code is located in tests and can test environment dynamics and RL agent features by pytest -v

After the setup, you can run experiment as follow (expected run in docker-container)

# python scripts/train_rl.py # train RL agent in grid environment
# python scripts/train_rl.py env.is_diff_drive=True # train RL agent in diff drive environmnet
# python scripts/train_rl.py env.is_discrete=False # train RL agent in continous environmnet
# python scripts/train_rl.py env.num_agents=10 # train RL agent in grid environment with 10 agent

This project builds upon or incorporates code and ideas from jaxmapp, by Ryo Yonetani and Keisuke Okumura:

Description

• Some parts of our implementation in the navigation environment are based on jaxmapp.
• Files that are implemented based on jaxmapp have it explicitly stated in their Docstring that they refer to jaxmapp.

Modification

• jaxmapp is primarily designed for the path planning task, while our repository focuses on the navigation task.
• The main differences are as follows:
  • We have adjusted the original implementation from jaxmapp to be more navigation-focused due to the differences in the intended tasks.
  • Added several codes suitable for reinforcement learning applications.

For additional details, please refer to jaxmapp.