Rofunc: The Full Process Python Package for Robot Learning from Demonstration

Rofunc package focuses on the robotic Imitation Learning (IL) and Learning from Demonstration (LfD) fields and provides valuable and convenient python functions for robotics, including demonstration collection, data pre-processing, LfD algorithms, planning, and control methods. We also provide an Isaac Gym-based robot simulator for evaluation. This package aims to advance the field by building a full-process toolkit and validation platform that simplifies and standardizes the process of demonstration data collection, processing, learning, and its deployment on robots.

Installation

Install from PyPI

The installation is very easy,

pip install rofunc

and as you'll find later, it's easy to use as well!

import rofunc as rf

Thus, have fun in the robotics world!

Note: Several requirements need to be installed before using the package. Please refer to the installation guide for more details.

Install from Source (Recommended)

git clone https://github.com/Skylark0924/Rofunc.git
cd Rofunc
pip install -r requirements.txt
pip install -e .

Besides, you need to install ZED SDK manually. (We have tried to package it as a .whl file to add it to requirements.txt, unfortunately, the ZED SDK is not very friendly and doesn't support direct installation.)

Documentation

Currently, we provide a simple document; please refer to here. A comprehensive one with both English and Chinese versions is built via the readthedoc. We provide a simple but interesting example: learning to play Taichi by learning from human demonstration.

To give you a quick overview of the pipeline of rofunc, we provide an interesting example of learning to play Taichi from human demonstration. You can find it in the Quick start section of the documentation.

The available functions and plans can be found as follows.

Classes	Types	Functions	Description	Status
Demonstration collection and pre-processing	Xsens	`xsens.record`	Record the human motion via network streaming	✅
		`xsens.process`	Decode the `.mvnx` file	✅
		`xsens.visualize`	Show or save gif about the motion	✅
	Optitrack	`optitrack.record`	Record the motion of markers via network streaming
		`optitrack.process`	Process the output `.csv` data	✅
		`optitrack.visualize`	Show or save gif about the motion
	ZED	`zed.record`	Record with multiple (1~n) cameras	✅
		`zed.playback`	Playback the recording and save snapshots	✅
		`zed.export`	Export the recording to mp4 or image sequences	✅
	Delsys EMG	`emg.record`	Record real-time EMG data via network streaming	✅
		`emg.process`	Filtering the EMG data	✅
		`emg.visualize`	Some visualization functions for EMG data	✅
	Multimodal	`mmodal.record`	Record multi-modal demonstration data simultaneously
		`mmodal.export`	Export multi-modal demonstration data in one line	✅
Learning from Demonstration	Machine learning	`dmp.uni`	DMP for uni-manual robot with several (or one) demonstrated trajectories
		`gmr.uni`	GMR for uni-manual robot with several (or one) demonstrated trajectories	✅
		`gmm.uni`	GMM for uni-manual robot with several (or one) demonstrated trajectories
		`tpgmm.uni`	TP-GMM for uni-manual robot with several (or one) demonstrated trajectories	✅
		`tpgmm.bi`	TP-GMM for bimanual robot with coordination learned from demonstration	✅
		`tpgmr.uni`	TP-GMR for uni-manual robot with several (or one) demonstrated trajectories	✅
		`tpgmr.bi`	TP-GMR for bimanual robot with coordination learned from demonstration	✅
	Deep learning	`bco`	Behavior cloning from observation	✅
		`strans`	Structured-Transformer method proposed in IEEE RAL
	Reinforcement learning	`ppo`	Proximal Policy Optimization (PPO) baseline
		`dqn`	Deep Q Network baseline
		`sac`	Soft Actor Critic baseline
		`cql`	Conservative Q-learning for fully offline learning
		`mcql`	Mixed conservative Q-learning for learning from demonstration with interaction
Planning	LQT	`lqt.uni`	Linear Quadratic Tracking (LQT) for uni-manual robot with several via-points	✅
		`lqt.bi`	LQT for bimanual robot with coordination constraints	✅
		`lqt.uni_fb`	Generate smooth trajectories with feedback	✅
		`lqt.uni_cp`	LQT with control primitive	✅
	iLQR	`ilqr.uni`	Iterative Linear Quadratic Regulator (iLQR) for uni-manual robot with several via-points	✅
		`ilqr.bi`	iLQR for bimanual robots with several via-points	✅
		`ilqr.uni_fb`	iLQR with feedback
		`ilqr.uni_cp`	iLQR with control primitive	✅
		`ilqr.uni_obstacle`	iLQR with obstacle avoidance	✅
		`ilqr.uni_dyna`	iLQR with dynamics and force control	✅
	MPC	`mpc.uni`	Model Predictive Control (MPC)
	CIO	`cio.whole`	Contact-invariant Optimization (CIO)
Tools	Logger	`logger.write`	General logger based on `tensorboard`
	Config	`config.get_config`	General config API based on `hydra`	✅
	VisuaLab	`visualab.trajectory`	2-dim/3-dim/with ori trajectory visualization	✅
		`visualab.distribution`	2-dim/3-dim distribution visualization	✅
		`visualab.ellipsoid`	2-dim/3-dim ellipsoid visualization	✅
	RoboLab	`robolab.transform`	Useful functions about coordinate transformation	✅
		`robolab.fk`	Forward kinematics w.r.t URDF file
		`robolab.ik`	Inverse kinematics w.r.t URDF file
		`robolab.fd`	Forward dynamics w.r.t URDF file
		`robolab.id`	Inverse dynamics w.r.t URDF file
Simulator	Franka	`franka.sim`	Execute specific trajectory via single Franka Panda arm in Isaac Gym	✅
	DualFranka	`dualfranka.sim`	Execute specific trajectory via dual Franka Panda arm in Isaac Gym
	CURI	`curi.sim`	Execute specific trajectory via human-like CURI robot in Isaac Gym	✅
	Walker	`walker.sim`	Execute specific trajectory via UBTECH Walker robot in Isaac Gym

Cite

If you use rofunc in a scientific publication, we would appreciate citations to the following paper:

@misc{Rofunc2022,
      author = {Liu, Junjia and Li, Zhihao and Li, Chenzui},
      title = {Rofunc: The full process python package for robot learning from demonstration},
      year = {2022},
      publisher = {GitHub},
      journal = {GitHub repository},
      howpublished = {\url{https://github.com/Skylark0924/Rofunc}},
}

The Team

Rofunc is developed and maintained by the CLOVER Lab (Collaborative and Versatile Robot Laboratory), CUHK.