gym-worm

OpenAI Gym environments for simulating the nematode C. elegans

Prerequisites

• python==3.8
• Swimmer-v3 requires mujoco-py dependencies; Swimmer-v4 requires mujoco dependencies
• install worm-assets and its dependencies

Dependencies for mujoco-py

• visit http://roboti.us/ and download MuJoCo 2.0 mujoco200_linux.zip, unzip it and get mujoco200_linux/, rename it to mujoco200/ and put it in ~/.mujoco folder
• download activation key mjkey.txt and put it in ~/.mujoco folder
• add environment variables in ~/.bashrc file, and run source ~/.bashrc

  export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/ubuntu/.mujoco/mujoco200/bin
  export MUJOCO_PY_MUJOCO_PATH=/home/ubuntu/.mujoco/mujoco200
  

• install mujoco-py: pip install mujoco-py==2.0.2.13
• install gym: pip install gym==0.22.0

Troubleshooting

Many errors may spring up while installing mujoco-py dependencies.

• run sudo apt update and sudo apt-get update first, and try the following
• sudo apt install gcc
• sudo apt install libosmesa6-dev libgl1-mesa-glx libglfw3
• sudo apt install libgl1-mesa-glx --reinstall
• sudo apt install libgl1-mesa-dev
• sudo apt-get install libc-dev
• conda install patchelf or sudo apt-get install patchelf

Dependencies for mujoco

• install mujoco: pip install mujoco
• install gym: pip install gym
• install imageio: pip install imageio
• install moviepy: pip install moviepy
• make sure that mujoco-py is NOT installed

Installation

• install the project in editable mode from local project path pip install -e .

Gym Debugging

Add the following lines in the script, it'll output info messages.

import gym
gym.logger.set_level(gym.logger.DEBUG)

Gym agents

• Capsule_v0: Capsule-v0 models skating effect with anisotropic friction
• ellipsoid2d_v0: Ellipsoid2D-v0 with only z-axis rotation dof
• ellipsoid_v0: Ellipsoid-v0 with specific n_bodies, joint_range and meshes
• ellipsoid_v1: Ellipsoid-v1 with specific n_bodies, joint_range, meshes and y-axis rotational degree of freedom
• head: head setup enabling free rotation at head and neck
• joint: remove joints and its corresponding actuators
• muscle_ellipsoid2d_v0: ellipsoid swimmer with tendon muscles and only z-axis rotation dof
• muscle_ellipsoid_v0: ellipsoid swimmer with tendon muscles
• muscle_swimmer_v0: muscle swimmer with specific n_bodies, joint_range, body_len and muscle_len
• muscle_worm_v0: muscle swimmer with specific joint_range, body_len, muscle_len and arrangement
• position_actuator: position actuators
• sphere_v0: Sphere-v0 with specific n_bodies, joint_range and meshes
• sphere_v1: Sphere-v1 with specific n_bodies, joint_range, meshes and y-axis rotational degree of freedom
• swimmer_v3: Swimmer-v3 modified to allow loading model from xml_str
• swimmer_v3_v1: Swimmer-v3 with specific n_bodies, joint_range and body_len
• swimmer_v4: Swimmer-v4 modified to allow loading model from xml_str
• trapped: trap specific bodies

Gym environments

• agar: agar environment containing composite object as agar based NGM
• camera: camera setup
• compositional_env: child models can be attached to parent models in order to create compositional scenes
• chemotaxis: chemotaxis environment containing a cylinder as chemical source
• forage: forage environment containing cylinders as chemical sources and box geoms as perimeter
• frame_sensor: frame sensors for framepos and framequat
• friction: friction environment containing box geoms as sections of floor with different friction setting
• gravity: gravity setup
• maze: maze environment containing box geoms as perimeter and walls
• pool: pool environment for simulating swimming behavior in 3D space
• position: position sensors of 3D center of mass and position
• sensor: actuator sensors
• tendon: tendon sensors of tendon length and velocity
• terrain: terrain environment
• touch: touch sensors at the head and tail
• tracking: tracking environment containing a controllable sphere as chemical source

Gym wrappers

• distribution_observation: wrapper for environmental substance distribution
• muscle_action: wrapper for decoding locomotion outputs of body wall muscles into control signals
• recorder: wrapper for selecting camera
• sensor_observation: wrapper for sensor data
• skip_frame: wrapper for skipping frames

Gym simulations

• circle: circular action for arena and no action for robot
• random: random action
• zero: no action

MuJoCo skin

• skin format: https://mujoco.readthedocs.io/en/latest/XMLreference.html#skin
• install early version of dm-control that supports MuJoCo 2.0

  pip install dm-control==0.0.364896371
  

• parse and serialize MuJoCo skin files with dm_control.mjcf.skin

References

• Gym: https://github.com/openai/gym
• how to create new environments for Gym: https://github.com/openai/gym/blob/master/docs/creating_environments.md
• third party environments: https://github.com/openai/gym/blob/master/docs/third_party_environments.md