README.md

metadata

license: apache-2.0
task_categories:
  - reinforcement-learning
  - other
language:
  - code
tags:
  - crafter
  - craftax
  - expert-demonstrations
  - skill-segmentation
  - action-segmentation
  - object-centric
pretty_name: Craftax Skill Segmentation Dataset
size_categories:
  - 1K<n<10K

Dataset Card for Craftax Expert Skill Data

This dataset consists of expert demonstration trajectories from the Crafter (Craftax) environment. Each trajectory includes ground-truth skill segmentation annotations, enabling research into action segmentation, skill discovery, imitation learning, and reinforcement learning with temporally-structured data.

Dataset Details

Dataset Description

The Craftax Skill Segmentation Dataset contains gameplay trajectories from an expert policy in the Crafter environment. Each trajectory is labeled with ground-truth skill boundaries and skill identifiers, allowing users to train and evaluate models for temporal segmentation, behavior cloning, and skill-based representation learning.

• Curated by: dami2106
• License: Apache 2.0
• Language(s) (NLP): Not applicable (code/visual)

Dataset Sources

• Repository: https://huggingface.co/datasets/dami2106/Craftax-Skill-Data
• Environment: CraftAX

Uses

Direct Use

This dataset is designed for use in:

• Training models to segment long-horizon behaviors into reusable skills.
• Evaluating action segmentation or hierarchical RL approaches.
• Studying object-centric or spatially grounded RL methods.
• Pretraining representations from visual expert data.

Out-of-Scope Use

• Language-based tasks (no natural language data is included).
• Real-world robotics (simulation-only data).
• Tasks requiring raw image pixels if they are not included in your setup.

Dataset Structure

Each data file includes:

• A sequence of states (top-down pixels, symbolic, pixel-obs, or PCA features).
• Corresponding actions (no-op action appended for final state).
• Skill labels marking where each skill begins and ends.

Example structure: Example structure:

{
  "task_id": "name of the task",
  "pixel_obs": [...],         // Raw visual observations (e.g., RGB frames)
  "top_down_obs": [...],      // Environment state from a top-down view
  "pca_features": [...],      // Compressed feature vectors 
  "actions": [...],           // Agent actions
  "groundTruth": [...],       // Ground-truth skill segmentation labels
  "mapping":                // Mapping metadata for skill ID -> groundTruth

}

Dataset Creation

Curation Rationale

This dataset was created to support research in skill discovery and temporal abstraction in visual reinforcement learning environments. The Craftax environment provides meaningful high-level tasks and object interactions, making it a suitable benchmark.

Source Data

Data Collection and Processing

• The expert trajectories were generated using a scripted or trained expert policy in the Craftax environment.
• Skill labels were annotated using environment signals (e.g., task success or inventory changes) and manual rules.

Who are the source data producers?

The data was generated programmatically in the Craftax simulation environment by an expert agent.

Annotations

Annotation process

Skill annotations were added using heuristics based on inventory changes, environment triggers, and task events. They were verified for consistency across episodes.

Who are the annotators?

Automated heuristics with manual inspection during development.

Bias, Risks, and Limitations

• The dataset is based on simulation, so real-world transferability may be limited.
• Skill labels are heuristically defined, which may not reflect a true underlying skill taxonomy.
• The expert behavior might be biased toward one specific strategy or task order.

Recommendations

Researchers should consider validating learned skills on diverse evaluation tasks. Skill segmentation boundaries are approximations and might not generalize well to different agents or environments.