MP_PyTorch: The Movement Primitives Package in PyTorch

MP_PyTorch package focus on Movement Primitives(MPs) on Imitation Learning(IL) and Reinforcement Learning(RL) and provides convenient movement primitives interface implemented by PyTorch, including DMPs, ProMPs and ProDMPs. Users can also implement custom Movement Primitives according to the basis and phase generator. Further, advanced NN-based Movement Primitives Algorithm can also be realized according to the convenient PyTorch-based Interface. This package aims to building a movement primitives toolkit which could be combined with modern imitation learning and reinforcement learning algorithm.

Installation

For the installation we recommend you set up a conda environment or venv beforehand.

This package will automatically install the following dependencies: addict, numpy, pytorch and matplotlib.

1. Install from Conda (Recommended)

conda install -c conda-forge mp_pytorch

2. Install from PyPI

pip install mp_pytorch

3. Install from source

git clone git@github.com:ALRhub/MP_PyTorch.git
cd mp_pytorch
pip install -e .

After installation, you can import the package easily.

import mp_pytorch
from mp_pytorch import MPFactory

Quickstart

For further information, please refer to the User Guide.

The main steps to create ProDMPs instance and generate trajectories are as follows:

1. Edit configuration

Suppose you have edited the required configuration. You can view the demo and check how to edit the configuration in Edit Configuration.

# config, times, params, params_L, init_time, init_pos, init_vel, demos = get_mp_utils("prodmp", True, True)

2. Initial prodmp instance and update inputs

mp=MPFactory.init_mp(**config)
mp.update_inputs(times=times, params=params, params_L=params_L,
init_time=init_time, init_pos=init_pos, init_vel=init_vel)
# you can also choose to learn parameters from demonstrations.params_dict=mp.learn_mp_params_from_trajs(times, demos)

3. Generate trajectories

traj_dict=mp.get_trajs(get_pos=True, get_pos_cov=True,
get_pos_std=True, get_vel=True,
get_vel_cov=True, get_vel_std=True)
# for probablistic movement primitives, you can also choose to sample trajectoriessamples, samples_vel=mp.sample_trajectories(num_smp=10)

The structure of this package can be seen as follows:

Types	Classes	Description
Phase Generator	PhaseGenerator	Interface for Phase Generators
	RhythmicPhaseGenerator	Rhythmic phase generator
	SmoothPhaseGenerator	Smooth phase generator
	LinearPhaseGenerator	Linear phase generator
	ExpDecayPhaseGenerator	Exponential decay phase generator
Basis Generator	BasisGenerator	Interface for Basis Generators
	RhythmicBasisGenerator	Rhythmic basis generator
	NormalizedRBFBasisGenerator	Normalized RBF basis generator
	ProDMPBasisGenerator	ProDMP basis generator
Movement Primitives	MPFactory	Create an MP instance given configuration
	MPInterface	Interface for Deterministic Movement Primitives
	ProbabilisticMPInterface	Interface for Probablistic Movement Primitives
	DMP	Dynamic Movement Primitives
	ProMP	Probablistic Movement Primitives
	ProDMP	Probablistic Dynamic Movement Primitives

Cite

If you interest this project and use it in a scientific publication, we would appreciate citations to the following information:

@article{li2023prodmp,
title={ProDMP: A Unified Perspective on Dynamic and Probabilistic Movement Primitives},
author={Li, Ge and Jin, Zeqi and Volpp, Michael and Otto, Fabian and Lioutikov, Rudolf and Neumann, Gerhard},
journal={IEEE Robotics and Automation Letters},
year={2023},
publisher={IEEE}
}

Team

MP_PyTorch is developed and maintained by the ALR-Lab(Autonomous Learning Robots Lab), KIT.

Welcome to our GitHub Pages!