jordykieto/ppo-hyrule-data

PPO Hyrule Data

_{More research: Learning in Dreams, Winning in Reality · Neural Simulators for Interactive Environments · Imitation Learning for Autonomous Driving in CARLA · all research ↗}

Combat trajectories from The Legend of PPO: Learning to Fight in Hyrule (GitHub, models): human and agent play of The Legend of Zelda: Ocarina of Time under the Dolphin emulator.

No game content is included. The data holds only numbers decoded from emulated memory: positions, velocities, health, timers and categorical actor states, plus actions and rewards. There are no frames, audio, ROM, savestates or memory dumps. The start-state savestates the environment needs are not redistributed. The GitHub repository has a savestate manifest and scripts/capture_boss_state.py, so owners of the game can create their own.

Contents

Folder	What	Episodes	Size	Release	Trained
human_gohma/	100 human victories over Queen Gohma: 54,626 steps at 10 Hz from 23 pre-fight starts, with raw pad input and camera yaw. MANIFEST.json gives the digests (v2 e57da6027a30…; v1 = the first 13 episodes).	100	43 MB	full	bc-human-100-strat (91 train / 9 val, event-stratified); the win class of rvs-m5; part of wm-obs-epoch120
gohma_ppo_sample/	PPO episodes against Gohma from 18 collector runs: all 452 victories and all 24 room exits, plus 300 timeouts and 1,000 deaths sampled at random (seed 0)	1,776 of 22,214	≈ 0.9 GB	sample, wins over-sampled	rvs-m5, wm-obs-epoch120 (see note)
deku_baba_v5_sample/	Deku Baba melee transitions with the full 363-d state: ranks 000–003, plus dataset.json and manifests	≈ 4.7k of 103,716	≈ 1.1 GB	sample	world-model and JEPA research (no released model)
deku_baba_v3/	the v3 Deku Baba sets (main, moves, approach, scenarios) with sjepa_prep_v3.json	398	≈ 16 MB	full	sjepa-p1
counterfactual_pairs/	150 pairs: the same state, two different move actions	—	57 KB	full	motion analysis
OBSERVATION_SCHEMA.md	every field, its offset source and units

How the samples relate to the full data.

• The full Deku Baba v5 set has 91.5 M transitions in 103,716 episodes (28 GB). The ranks were written by independent collectors, so ranks 000–003 are a representative slice.
• The full Gohma PPO set has 22,214 episodes in gohma_steps2, plus earlier collectors.
• The exact list of 700 PPO episodes behind rvs-m5 (200 wins + ~500 losses) was never pinned. This sample contains every win, so it covers that model's win class, but its loss episodes can only be approximated.

How the data is made

Dolphin 2603a runs the GameCube release of Ocarina of Time (NACE01) with the Null video backend, so nothing is rendered. The environment reads emulated RAM zero-copy and decodes named fields using offsets from the zeldaret decompilation:

• Raw state (state, 363-d): Link's 43 fields, plus 16 actor rows × 20 fields.
• Policy observation (obs): a featurised vector, 483-d (v4, Deku Baba) or 1,033-d (v7, Gohma).
• Actions: 63 or 108 discrete actions.
• Rate: 10 Hz.

fn_rel / action_fn_rel are offsets of the active state function, used as categorical state ids. They are numbers, not code. See OBSERVATION_SCHEMA.md.

Format

One self-describing .npz per episode. Keys vary by set:

• human: obs, action, reward, pad, cam_yaw, slot, obs_fields, action_space, recorded_utc, git_commit;
• PPO: obs, action, reward, slot, policy;
• v5: state, obs, action, reward.

The outcome is in each filename (_success, _death, _timeout, _room / _left_room, _WIN).

import numpy as np
ep = np.load("human_gohma/ep00_slot1_WIN.npz", allow_pickle=True)
print({k: ep[k].shape for k in ep.files if hasattr(ep[k], "shape")})

Citation

@techreport{kieto2026hyrule,
  title = {The Legend of {PPO}: Learning to Fight in Hyrule}, author = {Kieto, Jordy}, year = {2026},
  type = {Technical Report}, url = {https://github.com/JordyKieto/ppo-hyrule}}

The Legend of Zelda: Ocarina of Time is © Nintendo. This dataset is unofficial research data and is not affiliated with Nintendo.