README.md

---

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

Repository address: https://github.com/Skylark0924/Rofunc
Documentation: https://rofunc.readthedocs.io/

Rofunc package focuses on the Imitation Learning (IL), Reinforcement Learning (RL) and Learning from Demonstration (LfD) for (Humanoid) Robot Manipulation. It provides valuable and convenient python functions, including demonstration collection, data pre-processing, LfD algorithms, planning, and control methods. We also provide an IsaacGym and OmniIsaacGym 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.

• Rofunc: The Full Process Python Package for Robot Learning from Demonstration and Robot Manipulation
  • Citation
  • Update News 🎉🎉🎉
    • v0.0.2.6 Support dexterous grasping and human-humanoid robot skill transfer
  • Installation
  • Documentation
  • RofuncRL
  • Star History
  • Related Papers          
  • The Team
  • Acknowledge
    • Learning from Demonstration
    • Planning and Control

Citation

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

@article{liu2023rofunc,
	title={Rofunc: The Full Process Python Package for Robot Learning from Demonstration and Robot Manipulation},
	author={Liu, Junjia and Dong, Zhipeng and Li, Chenzui and Li, Zhihao and Yu, Minghao and Delehelle, Donatien and Chen, Fei},
	year={2023},
	journal={Zenodo, https://github.com/Skylark0924/Rofunc},
	doi={10.5281/zenodo.10016946},
}

Warning

If our code is found to be used in a published paper without proper citation, we reserve the right to address this issue formally by contacting the editor to report potential academic misconduct!

如果我们的代码被发现用于已发表的论文而没有被恰当引用，我们保留通过正式联系编辑报告潜在学术不端行为的权利。

Update News 🎉🎉🎉

• [2024-12-24] 🎮 Start trying to support Genesis simulator.

v0.0.2.6 Support dexterous grasping and human-humanoid robot skill transfer

• [2024-12-20] 🎉🚀 Human-level skill transfer from human to heterogeneous humanoid robots have been completed and are awaiting release. Preview
• [2024-01-24] 🚀 CURI Synergy-based Softhand grasping tasks are supported to be trained by RofuncRL.
• [2023-10-31] 🚀 RofuncRL: A modular easy-to-use Reinforcement Learning sub-package designed for Robot Learning tasks is released. It has been tested with simulators like OpenAIGym, IsaacGym, OmniIsaacGym (see example gallery), and also differentiable simulators like PlasticineLab and DiffCloth.
• ...
• If you want to know more about the update news, please refer to the changelog.

Installation

Please refer to the installation guide.

Documentation

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.

Note ✅: Achieved 🔃: Reformatting ⛔: TODO

Data		Learning		P&C		Tools		Simulator
xsens.record	✅	DMP	⛔	LQT	✅	config	✅	Franka	✅
xsens.export	✅	GMR	✅	LQTBi	✅	logger	✅	CURI	✅
xsens.visual	✅	TPGMM	✅	LQTFb	✅	datalab	✅	CURIMini	🔃
opti.record	✅	TPGMMBi	✅	LQTCP	✅	robolab.coord	✅	CURISoftHand	✅
opti.export	✅	TPGMM_RPCtl	✅	LQTCPDMP	✅	robolab.fk	✅	Walker	✅
opti.visual	✅	TPGMM_RPRepr	✅	LQR	✅	robolab.ik	✅	Gluon	🔃
zed.record	✅	TPGMR	✅	PoGLQRBi	✅	robolab.fd	⛔	Baxter	🔃
zed.export	✅	TPGMRBi	✅	iLQR	🔃	robolab.id	⛔	Sawyer	🔃
zed.visual	✅	TPHSMM	✅	iLQRBi	🔃	visualab.dist	✅	Humanoid	✅
emg.record	✅	RLBaseLine(SKRL)	✅	iLQRFb	🔃	visualab.ellip	✅	Multi-Robot	✅
emg.export	✅	RLBaseLine(RLlib)	✅	iLQRCP	🔃	visualab.traj	✅
mmodal.record	⛔	RLBaseLine(ElegRL)	✅	iLQRDyna	🔃	oslab.dir_proc	✅
mmodal.sync	✅	BCO(RofuncIL)	🔃	iLQRObs	🔃	oslab.file_proc	✅
		BC-Z(RofuncIL)	⛔	MPC	⛔	oslab.internet	✅
		STrans(RofuncIL)	⛔	RMP	⛔	oslab.path	✅
		RT-1(RofuncIL)	⛔
		A2C(RofuncRL)	✅
		PPO(RofuncRL)	✅
		SAC(RofuncRL)	✅
		TD3(RofuncRL)	✅
		CQL(RofuncRL)	⛔
		TD3BC(RofuncRL)	⛔
		DTrans(RofuncRL)	✅
		EDAC(RofuncRL)	⛔
		AMP(RofuncRL)	✅
		ASE(RofuncRL)	✅
		ODTrans(RofuncRL)	⛔

RofuncRL

RofuncRL is one of the most important sub-packages of Rofunc. It is a modular easy-to-use Reinforcement Learning sub-package designed for Robot Learning tasks. It has been tested with simulators like OpenAIGym, IsaacGym, OmniIsaacGym (see example gallery), and also differentiable simulators like PlasticineLab and DiffCloth. Here is a list of robot tasks trained by RofuncRL:

Note
You can customize your own project based on RofuncRL by following the RofuncRL customize tutorial.
We also provide a RofuncRL-based repository template to generate your own repository following the RofuncRL structure by one click.
For more details, please check the documentation for RofuncRL.

The list of all supported tasks.

Tasks	Animation	Performance	ModelZoo
Ant			✅
Cartpole
Franka Cabinet			✅
Franka CubeStack
CURI Cabinet			✅
CURI CabinetImage
CURI CabinetBimanual
CURIQbSoftHand SynergyGrasp			✅
Humanoid			✅
HumanoidAMP Backflip			✅
HumanoidAMP Walk			✅
HumanoidAMP Run			✅
HumanoidAMP Dance			✅
HumanoidAMP Hop			✅
HumanoidASE GetupSwordShield			✅
HumanoidASE PerturbSwordShield			✅
HumanoidASE HeadingSwordShield			✅
HumanoidASE LocationSwordShield			✅
HumanoidASE ReachSwordShield			✅
HumanoidASE StrikeSwordShield			✅
BiShadowHand BlockStack			✅
BiShadowHand BottleCap			✅
BiShadowHand CatchAbreast			✅
BiShadowHand CatchOver2Underarm			✅
BiShadowHand CatchUnderarm			✅
BiShadowHand DoorOpenInward			✅
BiShadowHand DoorOpenOutward			✅
BiShadowHand DoorCloseInward			✅
BiShadowHand DoorCloseOutward			✅
BiShadowHand GraspAndPlace			✅
BiShadowHand LiftUnderarm			✅
BiShadowHand HandOver			✅
BiShadowHand Pen			✅
BiShadowHand PointCloud
BiShadowHand PushBlock			✅
BiShadowHand ReOrientation			✅
BiShadowHand Scissors			✅
BiShadowHand SwingCup			✅
BiShadowHand Switch			✅
BiShadowHand TwoCatchUnderarm			✅

Star History

Related Papers          

1. Robot cooking with stir-fry: Bimanual non-prehensile manipulation of semi-fluid objects (IEEE RA-L 2022 | Code)

@article{liu2022robot,
         title={Robot cooking with stir-fry: Bimanual non-prehensile manipulation of semi-fluid objects},
         author={Liu, Junjia and Chen, Yiting and Dong, Zhipeng and Wang, Shixiong and Calinon, Sylvain and Li, Miao and Chen, Fei},
         journal={IEEE Robotics and Automation Letters},
         volume={7},
         number={2},
         pages={5159--5166},
         year={2022},
         publisher={IEEE}
}

2. SoftGPT: Learn Goal-oriented Soft Object Manipulation Skills by Generative Pre-trained Heterogeneous Graph Transformer (IROS 2023｜Code coming soon)

@inproceedings{liu2023softgpt,
               title={Softgpt: Learn goal-oriented soft object manipulation skills by generative pre-trained heterogeneous graph transformer},
               author={Liu, Junjia and Li, Zhihao and Lin, Wanyu and Calinon, Sylvain and Tan, Kay Chen and Chen, Fei},
               booktitle={2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
               pages={4920--4925},
               year={2023},
               organization={IEEE}
}

3. BiRP: Learning Robot Generalized Bimanual Coordination using Relative Parameterization Method on Human Demonstration (IEEE CDC 2023 | Code)

@inproceedings{liu2023birp,
               title={Birp: Learning robot generalized bimanual coordination using relative parameterization method on human demonstration},
               author={Liu, Junjia and Sim, Hengyi and Li, Chenzui and Tan, Kay Chen and Chen, Fei},
               booktitle={2023 62nd IEEE Conference on Decision and Control (CDC)},
               pages={8300--8305},
               year={2023},
               organization={IEEE}
}

The Team

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

Acknowledge

We would like to acknowledge the following projects:

Learning from Demonstration

1. pbdlib
2. Ray RLlib
3. ElegantRL
4. SKRL
5. DexterousHands

Planning and Control

1. Robotics codes from scratch (RCFS)