PokeDreamer Trajectory Dataset (v2)

This dataset contains high-resolution trajectories collected from playing Pokémon Red on the Game Boy using a PyBoy emulator. It is designed to train Recurrent State-Space Models (RSSM) and discrete latent dynamics models (such as DreamerV2/V3).

Dataset Summary

• Total Steps: 200,000 steps of transition pairs.
• Resolution: Native $160 \times 144$ pixels, 3 channels (RGB).
• Collection Policy: A mixture of stochastic trajectories using a frozen PPO agent (PWhiddy baseline) and random exploration over 5 starting save states (Pallet Town, Oak's Lab, Route 1, Viridian City, Pewter City).
• Environment Ticks: 24 ticks per action.

Data Structure

The dataset is saved in compressed .npz chunks (10,000 steps per chunk). Each chunk contains:

📸 Observation Fields

• obs: (N, 144, 160, 3) uint8 array of RGB frames.
• actions: (N,) uint8 array of action indices.
  • 0=down, 1=left, 2=right, 3=up, 4=a, 5=b, 6=start, 7=pass
• rewards: (N,) float32 array (placeholder 0.0).
• episode_starts: (N,) boolean array indicating the start of a trajectory segment.

🧠 WRAM State Fields (Metadata Probes)

• map_ids: (N,) uint8 array of map identifiers.
• xs: (N,) uint8 array of player X tile coordinates.
• ys: (N,) uint8 array of player Y tile coordinates.
• facings: (N,) uint8 array of facing directions (0=down, 4=up, 8=left, 12=right).
• in_battles: (N,) boolean array indicating if a battle is active.
• dialog_opens: (N,) boolean array indicating if a text box is open.
• badges: (N,) uint8 array of badge counts.
• party_hps: (N, 6) int16 array of current party member HPs.
• party_max_hps: (N, 6) int16 array of max party member HPs.

Usage

Load any chunk directly using NumPy:

import numpy as np

data = np.load("transitions_0001.npz")
frames = data["obs"]       # shape: (10000, 144, 160, 3)
actions = data["actions"]  # shape: (10000,)
print(f"Loaded chunk with {len(frames)} frames of shape {frames.shape[1:]}")