Status: Archive (code is provided as-is, no updates expected)

robogym is a simulation framework that uses OpenAI gym and MuJoCo physics simulator and provides a variety of robotics environments suited for robot learning in diverse settings.

This package has been tested on Mac OS Mojave, Catalina and Ubuntu 16.04 LTS, and is probably fine for most recent Mac and Linux operating systems.

Requires Python 3.7.4 or greater.

1. Install MuJoCo by following the instructions from the mujoco-py package.

2. To checkout the code and install it, via pip install, run:

   git clone [email protected]:openai/robogym.git
   cd robogym
   pip install -e .

   Or you can install it directly via:

   pip install git https://github.com/openai/robogym.git

Please use the below BibTeX entry to cite this framework:

@misc{robogym2020,
  author={OpenAI},
  title={{Robogym}},
  year={2020},
  howpublished="\url{https://github.com/openai/robogym}",
}

You can visualize and interact with an environment using robogym/scripts/examine.py.

For example, following scripts visualize the dactyl/locked.py environment.

python robogym/scripts/examine.py robogym/envs/dactyl/locked.py constants='@{"randomize": True}'

Note that constants='@{"randomize": True} is an argument to set constants for the environment.

Similarly, you can set parameters of an environment as well. Below shows a command for visualizing a block rearrange environment with 5 objects.

python robogym/scripts/examine.py robogym/envs/rearrange/blocks.py parameters='@{"simulation_params": {"num_objects": 5}}'

We support teleoperation for the rearrange environments via the --teleoperate option, which allows users to interact with an environment by controlling the robot with a keyboard. Below is an example command for the teleoperation.

python robogym/scripts/examine.py robogym/envs/rearrange/blocks.py parameters='@{"simulation_params": {"num_objects": 5}}' --teleoperate

Hold-out environments that are specified via a jsonnet config can also be visualized and teleoperated using this mechanism as below

python robogym/scripts/examine.py robogym/envs/rearrange/holdouts/configs/rainbow.jsonnet  --teleoperate

The environments extend OpenAI gym and support the reinforcement learning interface offered by gym, including step, reset, render and observe methods.

All environment implementations are under the robogym.envs module and can be instantiated by calling the make_env function. For example, the following code snippet creates a default locked cube environment:

from robogym.envs.dactyl.locked import make_env
env = make_env()

See the section on customization for details on how to customize an environment.

All the environment classes are subclasses of robogym.robot_env.RobotEnv. The classmethod RobotEnv.build is the main entry point for constructing an environment object, pointed by make_env in each environment. Customized parameters and constants should be defined by subclasses of RobotEnvParameters and RobotEnvConstants.

The physics and simulator setup are wrapped within robogym.mujoco.simulation_interface.SimulationInterface. There is a 1-1 mapping between one instance of SimulationInterface and one instance of RobotEnv.

Each environment contains a robot object accessible via env.robot that implements the RobotInterface.

Dactyl environments utilize a Shadow Robot hand robot simulation with 20 actuated degrees of freedom to perform in-hand manipulation tasks. Below is a full list of environments provided in this category:

Image	Name	Description
	dactyl/locked.py	Manipulate a locked cube with no internal degrees of freedom to match a target pose
	dactyl/face_perpendicular.py	Manipulate a Rubik's cube with 2 internal degrees of freedom to match a target pose and face angles
	dactyl/full_perpendicular.py	Manipulate a Rubik's cube with full internal 6 degrees of freedom to match a target pose and face angles
	dactyl/reach.py	Reach task for fingertip target positions

These environments are based on a UR16e robot equipped with a RobotIQ 2f-85 gripper that is able to rearrange a variety of object distributions in a tabletop setting. Several different types of robot control modes are supported as detailed here.

Various goal generators are provided to enable different tasks such as stack, pick-and-place, reach and rearrange to be specified on a given object distribution. List of all rearrange environments and their configs are described in this document.

Below is a list of object distributions supported in this category:

Image	Name	Description
	rearrange/blocks.py	Samples blocks of different colors
	rearrange/ycb.py	Samples from YCB objects
	rearrange/composer.py	Samples objects that are composed of random meshes that are either basic geom shapes or random convex meshes (decomposed YCB objects)
	rearrange/mixture.py	Generates objects from a mixture of mesh object distributions (supports ycb/geom mesh datasets)

For rearrange environments, we also provide a variety of hold-out tasks that are typically used for evaluation purposes. The goal states of various hold-out environments can be seen in the image grid below.

Most robotics environments support customization by providing additional parameters via constant argument to make_env, you can find which constants are supported by each environment by looking into definition of <EnvName>Constants class which usually lives under the same file as make_env. Some commonly supported constants are:

• randomize: If true, some randomization will be applied to physics, actions and observations.
• mujoco_substeps: Number of substeps per step for mujoco simulation which can be used to balance between simulation accuracy and training speed.
• max_timesteps_per_goal: Max number of timesteps allowed to achieve each goal before timeout.

Similarly, there are parameters arguments, which can be customized together with constants. You can find which parameters are supported by each environment by looking into definition of <EnvName>Parameters.

Below is the default settings that we use to train most of the robotics environments:

env = make_env(
    constants={
        'randomize': True,
        'mujoco_substeps': 10,
        'max_timesteps_per_goal': 400
    },
    parameters={
        'n_random_initial_steps': 10,
    }
)

Robogym provides a way to intervene the environment parameters during training to support domain randomization and curriculum learning.

Below shows an example of intervening the number of objects for blocks (rearrange) environment. You can use this interface to define a curriculum over the number of objects:

from robogym.envs.rearrange.blocks import make_env

# Create an environment with the default number of objects: 5
env = make_env(
    parameters={
        'simulation_params': {
            'num_objects': 5,
            'max_num_objects': 8,
        }
    }
)

# Acquire number of objects parameter interface
param = env.unwrapped.randomization.get_parameter("parameters:num_objects")

# Set num_objects: 3 for the next episode
param.set_value(3)

# Reset to randomly generate an environment with `num_objects: 3`
obs = env.reset()

See the document on "Interface for Environment Randomization" for more details.

We provide a set of tools to help create a customized rearrange environment via teleoperation.

See the document on "Build New Rearrange Environments" for more details.