tsilva/gradlab-breakout-trajectories

Breakout trajectories with HUD masking

Ordered action-conditioned Breakout trajectories collected with GradLab on September 12–13, 2026. All earlier trajectories are preserved; new gameplay augments the snapshot. Identical RGB captures share one lossless WebP image inside Hugging Face Parquet.

Quantity	Value
Transitions	6,728,910
Captured frame occurrences	6,731,230
Unique images	5,924,824
Reused captures	806,406
Complete episodes	2,241
Incomplete prefixes	79

Cumulative reuse is 11.9801%. The original augmentation target was 20% reused captures across the latest 100 completed episodes, weighted by capture counts. This includes one initial image and one successor per transition; source references are not extra captures. Reuse is tested against the entire prior image inventory, including earlier sessions. The 100-episode window starts fresh in this collection session.

Collection stopped at the operator's request before the reuse target or disk limit. Final reuse across the latest 100 completed episodes was 14.8111%; excluding their initial images and first 500 transitions, reuse was 3.4799%.

The top 17 rows of each 210×160 RGB capture are blackened before hashing and storage. All remaining pixels are preserved exactly; the policy still sees its original observations. Image reuse does not measure hidden simulator-state coverage or establish world-model quality.

Configuration/split	Rows
frames/assets	5,924,824
transitions/train	5,385,620
transitions/heldout	1,343,290
episodes/train	1,856
episodes/heldout	464
sessions/metadata	4

Choose trajectory splits before joining images. The shared image pool is storage, not an independent training split.

Read a transition and its images

import json
import numpy as np
from datasets import load_dataset

repo = "tsilva/gradlab-breakout-trajectories"
frames = load_dataset(repo, "frames", split="assets")
steps = load_dataset(repo, "transitions", split="train", streaming=True)

# Keep only the integer lookup in RAM; image bytes remain in the Arrow dataset.
ids = np.asarray(frames["frame_id"], dtype=np.int64)
order = np.argsort(ids)
ordered_ids = ids[order]

def frame(frame_id):
    position = int(np.searchsorted(ordered_ids, frame_id))
    if position == len(ordered_ids) or ordered_ids[position] != frame_id:
        raise KeyError(frame_id)
    return np.asarray(frames[int(order[position])]["image"])

row = next(iter(steps))
x = frame(row["source_frame_id"])
a = json.loads(row["native_action_json"])
y = frame(row["successor_frame_id"])
print(x.shape, a, y.shape)  # (210, 160, 3), executed action, (210, 160, 3)

Frame IDs are identifiers, not zero-based row offsets. For repeatable experiments, pass the same immutable Hub commit as revision= to every load_dataset call.

Build sequences

Transitions are stored in (episode_id, step) order within each split. For a sequence of L consecutive transitions, the frame IDs are the first source_frame_id followed by all L successor IDs; pair them with the L executed actions. Require a single episode ID, consecutive steps, and previous.successor_frame_id == next.source_frame_id. Never cross an episode boundary. A four-frame history can expose ball motion; longer recurrent histories and held-out multi-step rollouts should be evaluated for the intended world model.

Use native_action_json (or the matching effective_action_json) to supervise dynamics. The policy's selected action can differ because of the recorded auto_serve override. Action values are 0 = button, 1 = right, and 2 = left.

Each transition also includes policy/native/task rewards, termination and truncation flags, action-selection mode, sampling temperature, override ID, and session ID. Nullable fields mean unavailable. record_json retains the original typed facts, including native state diagnostics; those post-step diagnostics must not be used as current-state inputs when they expose future information. Its format contains a tagged JSON structure plus base64-encoded NumPy arrays with explicit dtypes and shapes; GradLab's collector read_record decodes it without loading a policy.

Collection and provenance

• Environment: ROM-free Breakout-Atari2600-v0, provider env-breakoutatari2600-turbo-native==0.5.12.
• PPO checkpoint: 77,495,040 training steps; SHA-256 f03b9b756b11179887182abe1f58930799405b108163513cf9a274653badff99.
• Configured frame skip 2. Actual elapsed native frames are unavailable; intermediate native frames are not included.
• The new session uses 32 independently seeded environments, eight CPU worker processes for stepping and lossless encoding, and one CUDA policy with temperature-grouped inference. Earlier session settings remain in their provenance.
• Temperatures 0.75, 1.0, 1.25 have probabilities 0.2, 0.6, 0.2; blocks contain 256 decisions; schedule seed 47.
• Full-game collection has a 3,000-step episode cap. Complete means an environment boundary was reached, including configured truncation; it does not imply the game was solved.
• Every fifth episode is held out, with separate episode seed allocation. New seeds continue the previous dataset allocation.
• Collector and worker source hashes, source commit, checkpoint contract and exact session limits are in session provenance and dataset-summary.json.

The earlier snapshot had 48 incomplete prefixes: 32 ended at earlier limits and 16 at the operator's final stop. The latter retain their original recording_failure label. One derived episode length and capture counter were reconciled against already-saved transitions; no image or transition was changed. finalization.json preserves that earlier reconciliation. The new collector handles stop signals between complete batches.

Every original committed Parquet file was checked against its prior checksum. All unique images and transition chains passed full integrity validation. Compact upload tables were compared with their source rows, metadata and encoded image bytes. upload.json records snapshot file hashes. SQLite, superseded records, weights and operator logs are excluded.

This is trajectory collection evidence, not policy acceptance or a demonstrated world-model benchmark. Source software attribution is GradLab and its native Breakout provider. No separate dataset license has been assigned.