The mppi_torch from tud-airlab

mppi_torch's Introduction

A fast and modular MPPI implementation with Halton spline sampling

Credits

Much of the backbone of the mppi implementation is based on pytorch_mppi, however it has been modified with addition sampling modes (e.g. Halton spline sampling) and support for ancillary controllers (priors) by Corrado Pezzato, Elia Trevisan and Chadi Salmi.

Structure

The project is structured as follows:

examples/: Contains the point robot example.
mppi_torch/: Contains the implementation of the MPPI algorithm and its utilities.
pyproject.toml and poetry.lock: Configuration files for project dependencies.

Installation

To install the project, follow these steps:

# Clone the repository
git clone <repository-url>

# Navigate to the project directory
cd <project-directory>

# Install dependencies
poetry install

Usage

To run the point robot example:

poetry shell
cd examples/point_robot
python run.py

Contributing

Contributions are welcome. Please submit a pull request.

Cite

This repository was originally developed for our paper Sampling-based Model Predictive Control Leveraging Parallelizable Physics Simulations. You can find that code here. If relevant, consider citing:

@misc{pezzato2023samplingbased,
      title={Sampling-based Model Predictive Control Leveraging Parallelizable Physics Simulations}, 
      author={Corrado Pezzato and Chadi Salmi and Max Spahn and Elia Trevisan and Javier Alonso-Mora and Carlos Hernandez Corbato},
      year={2023},
      eprint={2307.09105},
      archivePrefix={arXiv},
      primaryClass={cs.RO}
}

We also recently added the possibility of biasing the sampling distribution with ancillary controllers (or priors), as in Biased-MPPI: Informing Sampling-Based Model Predictive Control by Fusing Ancillary Controllers. Examples of how to use prior controllers with this repository can be found here. If you use this feature, consider citing:

@ARTICLE{trevisan2024biased,
  author={Trevisan, Elia and Alonso-Mora, Javier},
  journal={IEEE Robotics and Automation Letters}, 
  title={Biased-MPPI: Informing Sampling-Based Model Predictive Control by Fusing Ancillary Controllers}, 
  year={2024},
  volume={9},
  number={6},
  pages={5871-5878},
  keywords={Costs;Planning;Monte Carlo methods;Mathematical models;Optimal control;Vehicle dynamics;Trajectory;Motion and path planning;optimization and optimal control;collision avoidance;sampling-based MPC;MPPI},
  doi={10.1109/LRA.2024.3397083}}

mppi_torch's People

Contributors

Stargazers

Watchers

mppi_torch's Issues

Allow for arbitrary ancillary controllers

Add the possibility of passing an ancillary controller object, which should probably create a tensor of input sequences to be sampled by the mppi controller.
See: https://arxiv.org/abs/2401.09241

Did the halton-spline mode work

Dear author,

Hello! I have a question for the part of "halton-spline", the code in mppi.py is shown as below.

elif self.mppi_mode == 'halton-spline':
    # shift command 1 time step
    saved_action = self.mean_action[-1]
    self.mean_action = torch.roll(self.mean_action, -1, dims=0)
    self.mean_action[-1] = saved_action
    cost_total = self._compute_total_cost_batch_halton()
      
    action = torch.clone(self.mean_action)

In this part, the final 'action' is given by the mean of 'saved_action', and it only has a 'shift' operation and no other operations.

Is the method work, could you please explain this part. Thanks for your help.

Best wishes

Recommend Projects

tud-airlab / mppi_torch Goto Github PK

mppi_torch's Introduction

Credits

Structure

Installation

Usage

Contributing

Cite

mppi_torch's People

Contributors

Stargazers

Watchers

Forkers

mppi_torch's Issues

Allow for arbitrary ancillary controllers

Did the halton-spline mode work

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent