ARCLE is a RL environment for training Abstraction and Reasoning Corpus and ARC-like datasets, built on Farama Gymnasium.

ARCLE implements several environments with various actions on editing grids of ARC, with object-oriented grid operations. Inspired by O2ARC Interface, O2ARC 3.0 - the game-like human interface collecting human ARC solutions, ARCLE contains this interface as a RL environment, as well as the testing interface given with ARC Dataset.

Our work is presented at CoLLAs 2024! This paper contains initial experiments equipping ARCLE with PPO-optimized pixel-wise transformer on small grid pairs, in addition to a detailed description of ARCLE. Check out our Paper.

arclegif statetrans

• Simon ARC Env (Dec 2023): another ARC RL environment
• RL Algorithms (May 2024): DreamerV3, PPO working on ARCLE environments

Python >= 3.8

pip install arcle

PyPI Project Page

• Env Changes

  • O2ARCv2Env is accessible via O2ARCEnv or O2ARCEnv-v2.
  • Unify type of observations into np.int8 (that were previously np.uint8)
  • Unify observation space into a dictionary of Box or MultiBinary to flatten observations properly by Gymnasium APIs
  • Remove over-conditioned (@abstractmethod) requirements of AbstractARCEnv. Now user must only override create_operations only.

• Added action wrappers (BBoxWrapper, PointWrapper) to change selection to bounding box selection or one-point selection.

• Loader now supports definable random number generator.

• Orginized examples

• Bug Fix

  • Fixed terminated is not emitted when max_trial is infinity (-1)

• Python 3.8 Support (minimum requirement of Gymnasium)

• Env Changes

  • Rename ARCEnv into RawARCEnv in this version.
  • Removed MiniARCEnv. Please use RawARCEnv with loader=MiniARCLoader() instead.
  • New ARCEnv added, consisting action space of ARC testing interface given along with the ARC Dataset.
  • States in every environments are fully observable. All state-related instance variables are now in the current_state dictionary.
    • All operations receives state and action, and it changes state in-place.
    • You can deepcopy the state and call env.transition(state_copied, action) to get next state without changing original state on O2ARCv2Env. env.transition will be replaced as separated utility function in the future.
  • n-trial mode added. You canset maximum trials when you call gym.make() by putting argument max_trial=num. Unlimited trial mode is available when it is set by -1 (default).
  • Customizable Submit operation. It is defined in each env class as a method, not in a separated module.
    • You can specify boolean option reset_on_submit in env.reset (default=False)

• Bug fix

  • FloodFill operation without selection case fixed
  • Paste operation out-of-bound case fixed
  • CopyI operation out-of-bound case fixed
  • Apply patch exception handling

• Bug fix
  • Default all-or-none reward now gives only when submitted. It affects to existing all environments.

• O2ARCv2Env-v0
  • Changed ResizeGrid action to CropGrid action
  • Forced FloodFill action to select only one pixel (otherwise, it is NoOP)
  • Now CopyI, CopyO, Paste regards black pixels (pixel value 0) as a background. It copies pixels where the value is nonzero and the pixel is selected.
  • action['selection'] can handle int (automatically casts into np.bool_)
• Several exceptions handling

• ArcEnv-v0 & MiniArcEnv-v0 has renamed to ARCEnv-v0 & MiniARCEnv-v0.
• Changed internal current_grid variable to grid & grid_dim
• Added O2ARCv2Env-v0 environment.
• Added O2ARC Actions
  • Uses selection and selected. Some internal variables should be initialized.
  • Color, Flood Fill(DFS Color)
  • Move, Rotate, Flip
  • Copy, Paste and Clipboard Features (For O2ARC)

• Change Images
• Fix minor issues

• Initial Build.
• ArcEnv-v0 & MiniArcEnv-v0 Launched.