README.md

---
license: mit
language:
- en
pretty_name: human_overcooked_dataset
---

# Human Macro-Trajectory Dataset

This archive contains the human trajectory data used for offline capability-inference evaluation and human-data plotting.

## Contents

```text
macro_trajectories/
  human_macro_trajectory_dataset.npy
  0001/
    trajectory.npy
    key_frames.npy
    macro_trajectory.npy
    metadata.json
  0002/
    ...
```

## File Descriptions
human_macro_trajectory_dataset.npy
Preprocessed dataset used directly by human_data_offline_eval.py. This is the main file needed for offline inference evaluation.

trajectory.npy
Full state trajectory recorded during human play. This contains the environment state at each low-level/micro timestep.

key_frames.npy
Indices into trajectory.npy identifying the timesteps that correspond to macro-action decision points.

macro_trajectory.npy
Subsampled trajectory containing only the states at macro-action decision points:
```text
macro_trajectory = trajectory[key_frames]
```
This is the per-trajectory format used by the diversity and trajectory-analysis plotting code.

metadata.json
Metadata for the trajectory, including the map, task, ground-truth action restrictions, timestamp, and agent-belief configuration.

## Micro vs Macro States
The Overcooked environment runs at two levels.
Micro states are the low-level environment states observed at every primitive timestep. These correspond to individual movement/action updates, such as moving up/down/left/right or waiting while a macro action is being executed. These are stored in trajectory.npy.
Macro states are the decision-point states where agents choose a new macro action, such as getting an ingredient, chopping, getting a plate, or delivering. These are extracted from the micro trajectory using key_frames.npy and stored in macro_trajectory.npy.
The capability-inference algorithms operate on macro-level trajectories, so most analyses use macro_trajectory.npy or the bundled human_macro_trajectory_dataset.npy.