AnyCar to Anywhere: Learning Universal Dynamics Model for Agile and Adaptive Mobility

[Website] [Arxiv] [Video]

Environment Setup

Important

We recommend Ubuntu >= 22.04 + Python >= 3.10 + CUDA >= 12.3. You can create a mamba (or conda) environment before proceeding.

1. Install ROS2 Humble on your machine

2. Clone this repo and cd to the folder

   git clone git@github.com:LeCAR-Lab/anycar.git
   cd anycar

3. Create a new mamba environment and activate it

   mamba create -n anycar python=3.10
   mamba activate anycar

4. Install python dependencies

   pip install -r requirements.txt

5. Add the path of this project folder to your environment path CAR_PATH

   export CAR_PATH=$(pwd)

6. Add Python Path and Interpreter for ROS2 Compatibility

   • get python path: which python
   • get python path to packages: python -c "import site; print(site.getsitepackages()[0])"

   export PYTHON_EXECUTABLE=PATH-OF-PYTHON
   export PYTHONPATH=PATH-OF-PYTHON-SITE-PACKAGES:$PYTHONPATH

7. Build the project

   colcon build

8. Source the environment

   source install/setup.bash 

9. Download Foxglove Studio and import the visualization config from misc/anycar-vis.json

Quick Start

1. run foxglove node:

ros2 launch foxglove_bridge foxglove_bridge_launch.xml

2. run car_sim node:

XLA_PYTHON_CLIENT_MEM_FRACTION=0.3 ros2 launch car_ros2 car_sim.launch.py

3. Check the visualization in Foxglove Studio localhost:8765

Expected to see:

For this example environment, we provide an example checkpoint. Please put folder under car_foundation/car_foundation/models/.

AnyCar Pipeline

• Data Collection

  Please refer to car_collect/README.md for data collection. Run scripts in car_collect folder to collect data, the data will be automatically saved to car_foundation/car_foundation/data

• Model Training

  Please refer to car_foundation/README.md for model training. Run scripts in car_foundation folder to train model, the model will be automatically saved to car_foundation/car_foundation/models

• Controller.

  Please refer to car_dynamics/README.md for first-principle based dynamics model and sampling-based MPC implementation.

• Deployment.

  The deployment pipeline (sim/real) is implemented using ROS2, please refer to car_ros2/README.md for more details.

• Hardware Setup

  Please refer to hardware/README.md for our car configurations and 3d models.

Citation

@misc{xiao2024anycaranywherelearninguniversal,
      title={AnyCar to Anywhere: Learning Universal Dynamics Model for Agile and Adaptive Mobility}, 
      author={Wenli Xiao and Haoru Xue and Tony Tao and Dvij Kalaria and John M. Dolan and Guanya Shi},
      year={2024},
      eprint={2409.15783},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2409.15783}, 
}