Skip to content

Collections of robotics environments geared towards benchmarking multi-task and meta reinforcement learning

License

Notifications You must be signed in to change notification settings

NTU-Peer-Learning/Metaworld

 
 

Repository files navigation

Metaworld

License

Forked from Farama-Foundation/metaworld.

Installation

export PYTHONPATH=/path/to/metaworld

Usage

from metaworld.envs import ALL_ENVIRONMENTS as ALL_ENV

print(ALL_ENV)

Dimensions

MetaWorld

  • Observation space: 39
    • current, previous: 18 x 2
      • gripper xyz: 3 (0, 1, 2)
      • gripper distance apart: 1
      • object1 xyz: 3
      • object1 quat: 4
      • object2 xyz: 3
      • object2 quat: 4
    • goal pos: 3
  • Action space: 4
    • delta xyz: 3
    • gripper position: 1 (7)

Sawyer

DOF = 7 Joints = 9

  • Observation space: 51
    • joint qpos cos: 9
    • joint qpos sin: 9
    • joint qvel: 9
    • end effector pos: 3
    • end effector quat: 4
    • object1 xyz: 3
    • object1 quat: 4
    • object2 xyz: 3
    • object2 quat: 4
    • goal pos: 3
  • Action space: 8
    • joint torque: 7
    • gripper torque: 1

Jaco

DOF = 7 Joints = 13

  • Observation space: 63
    • joint qpos cos: 13
    • joint qpos sin: 13
    • joint qvel: 13
    • end effector pos: 3
    • end effector quat: 4
    • object1 xyz: 3
    • object1 quat: 4
    • object2 xyz: 3
    • object2 quat: 4
    • goal pos: 3
  • Action space: 8
    • joint torque: 7
    • gripper torque: 1

Fetch

DOF = 7 Joints = 9

  • Observation space: 49
    • current, previous: 23 x 2
      • joint qpos: 9
      • object1 xyz: 3
      • object1 quat: 4
      • object2 xyz: 3
      • object2 quat: 4
    • goal pos: 3
  • Action space: 8
    • joint torque: 7
    • gripper torque: 1

Environments

  • reach
  • button-press-topdown
  • button-press
  • drawer-close
  • drawer-open
  • window-close
  • window-open
  • plate-slide
  • sweep

Sawyer Environments

Worked:

  • button-press-topdown

  • button-press-topdown-wall

  • button-press

  • button-press-wall

  • dial-turn: quite hard

  • door-open: crowded

  • door-lock

  • drawer-close

  • drawer-open

  • faucet-open

  • faucet-close

  • handle-press

  • handle-pull

  • reach

  • plate-slide

  • plate-slide-side

  • plate-slide-back

  • plate-slide-back-side

  • sweep

  • window-close

  • window-open

Failed:

  • coffee: XML qpos order
  • assembly: too hard
  • basketball: too hard
  • bin-picking: too hard
  • box-close: too hard
  • disassemble: too hard
  • door-close: conflict initial position

About

Collections of robotics environments geared towards benchmarking multi-task and meta reinforcement learning

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 86.7%
  • Jupyter Notebook 13.3%