We introduce myGym, a toolkit suitable for fast prototyping of neural networks in the area of robotic manipulation and navigation. Our toolbox is fully modular, so that you can train your network with different robots, in several environments and on various tasks. You can also create a curriculum of tasks with increasing complexity and test your network on them. We also included an automatic evaluation and benchmark tool for your developed model. We have pretained the Yolact network for visual recognition of all objects in the simulator, so that you can reward your networks based on visual sensors only.

We keep training the current state-of-the-art algorithms to provide baselines for the tasks in the toolbox. There is also a leaderboard showing algorithms with the best generalization capability, tested on the tasks in our basic curriculum. From version 2.0 it is possible to train multiple networks within one task and switch between them based on reward or adaptively. The number of neteworks is specified in config file.

Learn more about the toolbox in our documentation

myGym 3.7

The latest version introduces new features:

• new robots (Nico, Tiago, HSR)
• new workspaces (human collaborative, Tiago table, Nico table)
• new algorithms for multi-step training
• visualization of multiple tranings in one graph
• sim2real for Nico robot
• new compositional rewards

This is the last version of myGym compactible with Stable Baselines and Python 3.7. Next version will be based on TF2 and Torch and Python 3.10.

Advantages

• Separate modules for fast prototyping (task.py, reward.py, env.py)

• Pretrained vision for instance wise semantic segmentation

• Human-robot collaboration environments

From version 2.1:

• Multi-step tasks defined inside config file with customizable observations

• Multi-goal rewards for training long horizon tasks

• REAL robotic gripping based on friction or containment

• Multi-network training - three networks switching in Pick and rotate task

Installation

Clone the repository:

git clone https://github.com/incognite-lab/mygym.git

cd mygym

Create Python 3.7 conda env as follows (later Python versions does not support TF 0.15.5 neccesary for Stable baselines ):

conda env create -f environment.yml

conda activate mygym

Install myGym:

python setup.py develop

If you face troubles with mpi4py dependency install the lib:

sudo apt install libopenmpi-dev

If you want to use the pretrained visual modules, please download them first:

cd myGym sh download_vision.sh

If you want to use the pretrained baseline models, download them here:

cd myGym sh download_baselines.sh

Check, whether the toolbox works:

sh ./speed_checker.sh

If everything is correct, the toolbox will train for two minutes without GUI and then shows the test results (at least 30% success rate)

Overview

Environment	Gym-v0 is suitable both single-step and multi-step manipulation and navigation
Workspaces	Table, Collaborative table, Maze, Vertical maze, Drawer, Darts, Football, Fridge, Stairs, Baskets
Vision	Cartesians, RGB, Depth, Class, Centroid, Bounding Box, Semantic Mask, Latent Vector
Robots	9 robotic arms, 2 dualarms, humanoid
Robot actions	Absolute, Relative, Joints
Objects	54 objects in 5 categories
Tasks	Reach, Press, Switch, Turn, Push, Pick, Place, PicknPlace, Poke,MultiReach, MultiPNP
Randomizers	Light, Texture, Size, Camera position
Baselines	Tensorflow, Pytorch
Physics	Bullet, Mujoco deprecated from version 2.0

Test the environments prior training

You can visualize the virtual gym env prior to the training.

python test.py

There will be the default workspace activated.

EXPERIMENTAL - You can control the robot and gripper from keyboard (arrows and A and Z for third axis in caartesian), spawn object to test the task (WIP)

There are also visual outputs from the active cameras (both RGB and Depth):

Find more details about this function in the documentation

Training

Run the default training without specifying the parameters:

python train.py

The training will start with the GUI window and a standstill visualization. Wait until the first evaluation to check the progress:

There are more training tutorials in the documentation

Single-step tasks

Reach

python train.py --config ./configs/train_reach.json

Press

python train.py --config ./configs/train_press.json

For details see documentation

Switch

python train.py --config ./configs/train_switch.json

For details see documentation

Turn

python train.py --config ./configs/train_turn.json

For details see documentation

Push (WIP)

python train.py --config ./configs/train_push.json

Poke (WIP)

python train.py --config ./configs/train_poke.json

Pick and Place

python train.py --config ./configs/train_pnp_1n.json

Multi-step tasks

Multi reach

python train.py --config ./configs/train_reach_multitask.json

Multi PNP

python train.py --config ./configs/train_pnp_3n_multitask2.json

Build a tower

python train.py --config ./configs/train_pnp_3n_multitask4.json

Parametric Training

As myGym is modular, you can easily train with different robots:

python train.py --robot jaco

You can also change the workspace within the gym, the task or the goal object. If you want to store an ouput video, just add the record parameter:

python train.py --workspace collabtable --robot panda --task push --task_objects wrench --record 1

Learn more about the simulation parameters in the documentation

Parallel Training

We have developed scripts for parallel training to speed up this process. You can edit the desired parameter in train_parallel.py and run it:

python train_parallel.py

Pretrained models

You can use the test script for the visualization of pretrained models:

python test.py --config ./trained_models/yourmodel/train.json

It will load the pretrained model and test it in the task and workspace defined in the config file.

Evaluation

There is automatic evaluation and logging included in the train script. It is controlled by parameters --eval_freq and --eval_episodes. The log files are stored in the folder with the trained model and you can easily visualize the learning progress after the training. There are also gifs for each eval period stored to compare the robot performance during training. We have also implemented evaluation in tensorboard:

tensorboard --logdir ./trained_models/yourmodel

If you want to interactively compare different parameters, just run tensorboard without model dir specification:

There are also other visualization scripts (Documerntation in preparation)

Environment

As myGym allows curriculum learning, the workspaces and tasks are concentrated in single gym, so that you can easily transfer the robot. The basic environment is called Gym-v0. There are more gyms for navigation and multi-agent collaboration in preparation.

Robots

Robot	Type	Gripper	DOF	Parameter value
UR-3	arm	no gripper	6	ur3
UR-5	arm	no gripper	6	ur5
UR-10	arm	no gripper	6	ur10
Kuka IIWA	arm	magnetic, gripper	6	kuka
Reachy	arm	passive palm	7	reachy
Leachy	arm	passive palm	7	leachy
Franka-Emica	arm	gripper	7	panda
Jaco arm	arm	two finger	13	jaco
Gummiarm	arm	passive palm	13	gummi
Human Support Robot (HSR)	arm	gripper	7	hsr
ABB Yumi	dualarm	two finger	12	yumi
ReachyLeachy	dualarm	passive palms	14	reachy_and_leachy
Pepper	humanoid	--	20	WIP
Thiago	humanoid	--	19	WIP
Atlas	humanoid	--	28	WIP

Workspaces

Name	Type	Suitable tasks	Parameter value
Tabledesk	manipulation	Reach,Press, Switch, Turn, PicknPlace	table
Drawer	manipulation	Pick, Place, PicknPlace	drawer
Fridge	manipulation	Push, Pick	fridge
Baskets	manipulation	Throw, Hit	baskets
Darts	manipulation	Throw, Hit	darts
Football	manipulation	Throw, Hit	football
Collaborative table	collaboration	Give, Hold, Move together	collabtable
Vertical maze	planning	--	veticalmaze
Maze	navigation	--	maze
Stairs	navigation	--	stairs

Baselines

Workspace	Reach	Pick	Place	PicknPlace	Push	Press	Switch	Turn	Poke	Multistep PNP	Multistep Reach
Tabledesk	✔️	✔️	✔️	✔️	✔️	✔️	✔️	✔️	✔️	✔️	✔️
Drawer	✔️	✔️	✔️	✔️	✔️	✔️	✔️	✔️	✔️	✔️	✔️
Collaborative table	✔️	✔️	✔️	✔️	✔️	✔️	✔️	✔️	✔️	✔️	✔️

Learnability

The new global evaluation metric, which we call \textit{learnability}, allows the user to evaluate and compare algorithms in a more systematic fashion. Learnability is defined as a general ability to learn irrespective of environmental conditions. The goal is to test an algorithm with respect to the complexity of environment. We have decomposed the environment complexity into independent scales. The first scale is dedicated to the complexity of the task. Second scale exploits the complexity of the robotic body that is controlled by the neural network. The third scale stands for the temporal complexity of the environment.

Leaderboard

Learnability is represented as a single value metric that evaluates algorithms under various conditions, allowing us to compare different RL algorithms. The number of conditions is limited for practical reasons, as the number of training configurations grows exponentially with each new condition, and each configuration requires standalone training and evaluation. Therefore, we limited the total number of combinations to $3^3$ = 27, which can be evaluated in few hours with a standard computer infrastructure.

Pos.	Algorhitm	Score
1.	PPO2	30.11
2.	TRPO	28.75
3.	ACKTR	27.5
4.	SAC	27.43
5.	PPO	27.21
5.	myAlgo	15.00

Authors

Incognite lab - CIIRC CTU

Core team:

Michal Vavrecka

Gabriela Sejnova

Megi Mejdrechova

Nikita Sokovnin

Contributors:

Radoslav Skoviera, Peter Basar, Michael Tesar, Vojtech Pospisil, Jiri Kulisek, Anastasia Ostapenko, Sara Thu Nguyen

Citation

'@INPROCEEDINGS{9643210, author={Vavrecka, Michal and Sokovnin, Nikita and Mejdrechova, Megi and Sejnova, Gabriela},

booktitle={2021 IEEE 33rd International Conference on Tools with Artificial Intelligence (ICTAI)},

title={MyGym: Modular Toolkit for Visuomotor Robotic Tasks},

year={2021}, volume={}, number={}, pages={279-283},

doi={10.1109/ICTAI52525.2021.00046}}'

Paper

myGym: Modular Toolkit for Visuomotor Robotic Tasks