tabletask

A minimal format to author tabletop manipulation tasks in simulation, from simple to complex, as jsons.

Example:

{
  "id": "robolab-banana_in_bowl",
  "instruction": "Pick up the banana and place it in the bowl.",
  "start": [
    ["on", "banana", "table"],
    ["on", "bowl", "table"]
  ],
  "goal": [
    ["inside", "banana", "bowl"]
  ]
}

Install:

pip install tabletask

Tasks ship with the package; the assets they name are fetched from the library the first time a task asks for them and kept in ~/.cache/tabletask, so an install stays small. Set TABLETASK_ASSETS to keep them elsewhere, or TABLETASK_LIBRARY_URL to serve them from somewhere else. Working in a checkout, the assets beside you are used and nothing is fetched.

Usage:

import tabletask
from tabletask.robots import Droid

env = tabletask.make("tasks/base/robolab-banana_in_bowl.json").add(Droid())
obs, info = env.reset()
action = [*obs["robot"]["joint_position"], 0.0]
obs, reward, terminated, truncated, info = env.step(action)
env.close()

MuJoCo runs the physics. tabletask.make(task, monitor_interval=0.01) sets how often task conditions are checked, in seconds; you can also set env.monitor_interval before stepping. The interval is rounded to physics steps (minimum one) and does not change the physics timestep.

Acknowledgement:

Tasks ported from ABC and RoboLab.

Licence: MIT, for the library and the tasks. The ported assets carry the terms of where they came from; assets/sources/ holds their manifests and licence notices.