Welcome to myGym’s documentation!

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 you can train your network to control different robots in several envinronments defined parametrically. You can also create curicullum of tasks and test your network set of tasks with inreasing complexity. There is automatic evaluation and benchmark tool for your network. We pretrained the neural networks for visual recognition of all objects in the simulator. We constantly train networks to provide baselines for the tasks in the toolbox. The training is 50x faster with the visualization turned on than realtime simulations.

Note

Mygym is now under construction

Overview

Environment	Gym-v0 is suitable for manipulation, navigation and planning tasks
Workspaces	Tabledesk, Collaborative table, Maze, Vertical maze, Drawer, Darts, Football, Fridge, Stairs, Baskets
Vision	Cartesians, RGB, Depth, Class, Centroid, Bounding Box, Semantic Mask, Latent vector
Robots	7 robotic arms, 2 dualarms, humanoid
Robot actions	Absolute, Relative, Joints
Objects	54 objects in 5 categories
Tasks	Reach, Push, Pick, Place, PicknPlace, Throw, Hit, Catch, Navigate
Randomizers	Light, Texture, Size, Camera position
Baselines	Tensorflow, Pytorch

Modular Structure

We developed fully modular toolbox where user can easily combine the predefined elements into custom envinronment. There are specific modules for each component of the simulation. User can easily modify and add custom modules.

How to

• Install
• Visualize workspace
• Train a robot
• Set parameters
• Edit config file
• Train robot - supervised vision
• Train a robot - unsupervised vision
• Train in parallel
• Evaluate with tensorboard
• Test trained model

How to - advanced

• Create custom reward
• Create custom workspace
• Create custom network
• Generate dataset

Gym

• Workspaces
• Objects
• Robots
• GymEnv

Baselines

• Baseline networks

Important classes

• BaseEnv
• Robot
• Task
• Reward
• Vision
• Camera
• Object

Citing myGym

@misc{myGym,
  author = {},
  title = {myGym},
  year = {2020},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {},
}

Authors

Core team:

Michal Vavrecka, Gabriela Sejnova, Megi Mejdrechova, Nikita Sokovnin

Contributors:

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

Paper

myGym: Modular Toolkit for Visuomotor Robotic Tasks

Indices and tables

• Index

• Module Index

• Search Page