Request access to OriginLab Game Recordings v0.3.0
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Two access tracks (see LICENSE.md). Internal Evaluation License: 90-day
internal evaluation, no publication or release obligation, no deployment or production use; delete or
convert at the end. Full dataset / commercial: production training and deployment rights per agreement
with Origin Lab. No redistribution of the data in any form.
Log in or Sign Up to review the conditions and access this dataset content.
OriginLab Game Recordings v0.3.0
Human gameplay captured in-engine under per-title licenses at 1080p / 60 FPS CFR on one shared frame clock: every stream starts at frame 0 and frame k matches frame k across pre-HUD and post-HUD RGB, surface normals, metric depth, audio, camera telemetry, keyboard and mouse inputs, in-engine action events and game state, and world telemetry — plus per-frame training tables.
Watch full playable previews of every modality, side by side and in sync, at app.originlab.ai/data.
Four games, side by side. Each tile is one session's mosaic: its four visual streams composed as quadrants of a single decoded video, so within a tile the streams cannot drift.
| Sessions | 144 |
| Hours | 145.1 |
| Games | 10 |
| Size | ~4.9 TB |
| Capture | 1080p / 60 FPS CFR, in-engine SDK, shared frame clock |
Dataset summary
This dataset is human gameplay recorded from inside the engine rather than from a screen capture. Every session pairs what the player saw with what the engine computed: pre-HUD and post-HUD RGB, dense depth and surface normals for every frame, game audio, per-frame camera pose, the complete keyboard and mouse stream, and time-ranged labels for in-game events.
Two properties define the release. Alignment: every stream in a session shares the same t=0 — depth frame k is RGB frame k, telemetry timestamps are milliseconds on the same clock, and the frame index is the only join key required. No offsets are shipped because none exist. Coverage: capture is directed, not passive. Sessions run under a proprietary mission-direction process designed to elicit the widest, least redundant range of actions, environments, and world states per hour; fewer than 5% of sampled frames near-duplicate a frame from any other session (measured below).
Supported tasks
| Task | Why this dataset |
|---|---|
| Depth estimation | Dense engine-rendered relative depth for every RGB frame (log-encoded, comparable across time): no pseudo-labels, no stereo reconstruction. Ideal for scale-invariant / ordinal MDE. |
| World models & video prediction | Long, continuous sessions with dense action and camera conditioning signals. |
| Imitation learning | Frame-level keyboard/mouse actions paired with what the player saw. |
| Camera pose & ego-motion | Per-frame 6-DOF camera extrinsics straight from the engine. |
Games
Titles are anonymized for licensing; real titles are disclosed under the full-dataset agreement.
| Game | Sessions | Hours | Size (GB) |
|---|---|---|---|
| Game 1 | 20 | 20.1 | 686 |
| Game 2 | 16 | 16.1 | 560 |
| Game 3 | 11 | 11.0 | 378 |
| Game 4 | 17 | 17.1 | 578 |
| Game 5 | 17 | 17.1 | 573 |
| Game 6 | 13 | 13.1 | 442 |
| Game 7 | 10 | 10.3 | 347 |
| Game 8 | 11 | 11.1 | 387 |
| Game 9 | 11 | 11.0 | 384 |
| Game 10 | 18 | 18.1 | 609 |
Dataset composition
Beyond hours and titles, here is what the footage actually contains. We sampled frames from each session (9,214 across all 144 sessions) and embedded them with a vision-language model to profile scene type, terrain, and lighting. These labels are assigned zero-shot by image-text similarity, so read the percentages as an approximate composition - the ordering is reliable, the exact figures are not calibrated ground truth.
Near-duplicate frames are rare: fewer than 5% of sampled frames closely match a frame from a different session, consistent with a capture process built to spread coverage rather than repeat it.
Scene and terrain
The corpus splits roughly 57.5% outdoor / 42.5% indoor, across a wide range of terrain:
| Terrain | Share |
|---|---|
| Interior | 17% |
| Cave / Underground | 16.5% |
| Desert | 15.3% |
| Grassland / Open | 13.5% |
| Snow / Alpine | 11.6% |
| Industrial / Facility | 9% |
| Forest | 6.9% |
| Urban / Built | 5.1% |
| Water / Coast | 5% |
Time of day and weather
Lighting and weather are only meaningful outdoors, so they are measured over outdoor frames only - indoor scenes are excluded rather than mislabeled as "night":
| Condition (outdoor frames) | Share |
|---|---|
| Daylight | 50.5% |
| Night | 26% |
| Fog / Haze | 14.6% |
| Rain / Storm | 6.2% |
| Dusk / Dawn | 1.8% |
| Overcast | 1% |
Maps of the corpus
Each map places every sampled frame in a low-dimensional space to show how the corpus distributes across scene type. The first two use interpretable content axes - defined by the concept references themselves - so a frame's position is a direct read of its content.
Enclosure and lighting. Indoor <-> outdoor across, night <-> daylight up: enclosed dark scenes fall to the lower-left, open daylit terrain to the upper-right.
Terrain manifold. A concept-space projection lays the terrains out as a continuous manifold, with titles spread by content rather than by studio style.
Both maps project onto the concept axes, so they reflect how the corpus distributes across scene type - not a separate measure of visual diversity. For contrast, the raw embedding - its axes chosen by the model, not by us - shows how visually distinct the titles are from one another, colored four ways (title, terrain, indoor/outdoor, and outdoor weather):
Composition by title
The anonymized titles occupy distinctly different regions of scene space - from almost entirely enclosed to near-fully outdoor - which is what makes the corpus useful for cross-title work rather than ten variations of one look:
| Game | Sessions | Outdoor | Dominant terrain |
|---|---|---|---|
| Game 1 | 20 | 77.3% | snow / alpine, grassland / open, desert |
| Game 10 | 18 | 32.8% | cave / underground, interior, desert |
| Game 4 | 17 | 53.9% | interior, snow / alpine, desert |
| Game 5 | 17 | 78.8% | grassland / open, interior, desert |
| Game 2 | 16 | 49.9% | cave / underground, snow / alpine, desert |
| Game 6 | 13 | 6.7% | industrial / facility, interior, cave / underground |
| Game 3 | 11 | 78.1% | forest, cave / underground, water / coast |
| Game 8 | 11 | 89.6% | desert, snow / alpine, urban / built |
| Game 9 | 11 | 96.6% | grassland / open, desert, forest |
| Game 7 | 10 | 9.2% | interior, cave / underground, industrial / facility |
Method: frames sampled per session across 143 sessions, embedded with SigLIP; terrain and lighting assigned by zero-shot image-text similarity (approximate, not calibrated); weather computed over outdoor frames only; near-duplicate = cosine >= 0.95 to a frame from a different session.
Dataset structure
One folder per session, eighteen files grouped by modality (video/, depth/, telemetry/, tables/), no archives to unpack, plus telemetry/gameclock.jsonl where the title exposes a readable clock. A 1-hour session measures about 35 GB: four video renditions (31 GB), the depth stream (3.8 GB), and roughly 200 MB of telemetry and tables. Everything the capture recorded ships decomposed — there is no container to parse.
There are no predefined train/validation/test splits: the release ships as whole sessions, and metadata/sessions.parquet is the per-session index (anonymized game label, duration, sync status, byte counts) from which to cut your own.
Data files and fields
| File | Purpose |
|---|---|
video/prehud.mp4 |
1080p H.264, 60 FPS CFR, HUD removed, in-engine capture with game audio. No trim: every stream starts at the shared frame 0 |
video/posthud.mp4 |
The frame exactly as the player saw it, HUD included, on the same 60 FPS clock |
video/normals.mp4 |
Per-pixel surface orientation rendered by the engine (world frame), same clock |
video/mosaic.mp4 |
The four visual streams composed as quadrants of one video — a playable sync proof, not extra data |
video/mosaic_layout.json |
Which mosaic quadrant holds which stream |
depth/depth.hevc |
10-bit HEVC elementary stream, log-encoded relative depth; frame k is RGB frame k |
depth/depth_meta.jsonl |
Per-frame depth params: depth_transform, range, FOV |
depth/decode_contract.json |
Session decode contract: near plane (makes depth metric), pinhole intrinsics, per-stream frame accounting (fps, dup counts), and the world frame |
telemetry/camera.jsonl |
Camera pose keyed by frame index and QPC (~2 samples per frame): position, pitch/yaw/roll |
telemetry/input.jsonl |
Mouse (position and dx / dy deltas), keyboard (key_down / key_up with modifiers), scroll, and window-focus events, frame-indexed |
telemetry/events.jsonl |
In-engine action events: a schema kind (weapon_fired, enemy_killed, item_pickup, ...) plus the engine's own label, frame-indexed |
telemetry/state.jsonl |
Sampled game state (health, active tool, ...): field, unit, and value per sample, frame-indexed |
telemetry/annotation.jsonl |
AI mechanic-detection intervals (label, category, confidence) where available |
telemetry/world.json |
World telemetry: engine, world-to-meters scale, handedness, gravity |
telemetry/gameclock.jsonl |
In-game clock samples, where the title exposes a readable clock (roughly two thirds of sessions) |
tables/frames.parquet |
One row per video frame, training-ready: camera pose, held-keys bitmask, mouse deltas, state columns, and event flags, all pre-joined on the frame index |
tables/events.parquet |
One row per in-engine event with resolved labels |
tables/conversion_manifest.json |
Every binning rule and cap used to build the tables, so they are re-runnable |
session.json |
Manifest: files + sizes, fps, the shared-clock alignment statement, frame accounting, and the sync audit (sync_status / sync_report) |
In-engine events & game state
This release adds the game's own event stream, recorded in-engine on the shared frame clock. Action events carry a stable schema kind plus the engine's literal label, so "the player fired" and "which engine function said so" are both on the frame where it happened. Game state is sampled continuously; the tables pre-join both onto the per-frame grid.
// telemetry/events.jsonl: in-engine action events (frame on the shared clock)
// (illustrative records: labels carry the engine's literal function name)
{"frame":203057,"qpc":1200000000001,"kind":"weapon_fired",
"source":"ue_processevent","confidence":255,"unit":"raw",
"arg_i":0,"arg_f":0.0,"label":"OnPrimaryFireShot"}
{"frame":203399,"qpc":1200000005702,"kind":"enemy_killed",
"source":"ue_processevent","confidence":255,"unit":"raw",
"arg_i":1,"arg_f":0.0,"label":"HandleTargetDeath"}
// telemetry/state.jsonl: sampled game state (continuous polling)
{"frame":0,"qpc":2333965391459,"field":"health",
"source":"memory_poll","unit":"normalized_0_1","value":1.0}
{"frame":8811,"qpc":2334112289031,"field":"active_tool",
"source":"memory_poll","unit":"raw","value":3}
Dataset creation
Consenting, compensated players record long sessions with our in-engine SDK, guided by a proprietary direction process: open-ended missions and challenges designed to draw out the widest, least redundant range of actions, environments, and world states a model can learn from. Depth and camera state are read from the engine at capture time, so depth is a measurement, not an estimate. Depth ships byte-identical to capture, never re-encoded, and every stream is aligned to the same shared clock before delivery. Before a session ships, an automated audit measures RGB-to-depth alignment across the whole session: a session certifies when the measured lag holds within a frame, and any session measured beyond that is withheld. Footage too static to measure ships marked unverified rather than certified, and sync_status in the metadata says which.
| Spec | Value |
|---|---|
| Resolution | 1920 x 1080 |
| Frame rate | 60 FPS, constant (CFR), one shared frame clock across every stream |
| RGB codec | H.264 high profile |
| Renditions | Pre-HUD RGB, post-HUD RGB, surface normals, and the composed mosaic ship as separate video files on the same clock; game audio rides in the video renditions |
| Depth | depth.hevc: dense 10-bit HEVC, log-encoded, planar Z. Metric centimeters = relative x near_plane_cm from the decode contract |
| Alignment | shared_clock_v1: every stream starts at frame 0 and frame k matches frame k — no trims, no offsets |
| Events | In-engine action events and sampled game state from the SDK event schema, frame-indexed |
| Audio | Game audio from the title process |
How to use it
You need ffmpeg and numpy, nothing else. RGB is a normal H.264 mp4; depth/depth.hevc is a plain HEVC stream you decode with ffmpeg -f hevc (gray16le), and the value is log-encoded relative depth (0 = near, 1/max = far; comparable across frames). For training, use the encoded value directly (or 1 − value as nearness): decoding to linear-ish depth crushes near-heavy scenes toward 0. Mask value == max (sky / far clip: undefined depth). Iterate both streams together; frame k of one is frame k of the other. tools/load_session.py in this repository shows the complete decode pattern; it was written against the flat v1 filenames, so point it at the video/, depth/, and telemetry/ paths from the table above.
# depth_transform = 2 (logarithmic), K = 4000
# luma: 0..65535 (ffmpeg gray16le) or 0..1023 (raw 10-bit); normalize to [0,1]
import numpy as np
K = 4000.0
def decode_depth(luma): # luma: uint array from the HEVC frame
q = luma.astype(np.float32) / 65535.0 # /1023.0 if you read raw 10-bit
d = (2.0 ** (q * np.log2(1.0 + K)) - 1.0) / K
return d # relative depth in [0,1]: 0 = NEAREST, 1 = farthest
valid = luma < 65535 # value == max => sky / far clip: undefined depth, mask out of losses
# METRIC: multiply by near_plane_cm from depth/decode_contract.json. This is planar Z
# (along the optical axis), not Euclidean ray distance.
What is in this repository
This gated repository carries the discovery surface of the corpus:
metadata/sessions.parquet— the full session index of the release (anonymized game labels, duration, sync status, byte counts). This is what the dataset viewer renders.metadata/files.parquet— the full-corpus file manifest (per-session file names and sizes), so you can audit a delivery against what this card advertises.assets/previews/— mosaic preview GIFs.tools/— the loader, depth decoder, and the parallel resumable downloader used for full-corpus delivery.
Pull a sample session with:
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="originlab/game-recordings-v3",
repo_type="dataset",
allow_patterns=["metadata/*", "sessions/932a7f81-763e-4e0e-826d-a466c213482f/*"],
local_dir="./v3",
)
Get the full corpus
The full release (144 sessions, 145.1 hours, ~4.9 TB) is delivered at scale rather than through browser downloads. Once your organization's access request is approved, you receive a download manifest: a signed list of every file in your selection with secure links valid for seven days. Feed it to tools/download_dataset.py (or any parallel download tool): the transfer runs multi-connection and resumable — an interrupted pull picks up where it left off, and already-complete files are skipped. On a 1 Gbit/s line a full session lands in minutes and a large selection overnight. Teams on AWS can request direct in-cloud access for the fastest possible transfer.
Request access with the form above, or contact Origin Lab directly. Interactive previews, per-game stats, and the release changelog live at app.originlab.ai/data.
Considerations for using the data
- Mechanic annotations are model-generated with confidence scores; coverage and label vocabulary vary by title.
- Depth comes from the engine's depth buffer, so translucent effects (fog, glass, particles) follow how the engine renders them.
- Sessions are guided free play: a proprietary direction process steers players through open-ended missions and challenges to maximize action diversity and minimize redundancy, so the action distribution is broader than natural play. Idle and menu time still occurs and is flagged in the annotations where detected.
- Depth range varies by genre and player behavior: many frames are near-field dominant and rarely reach the far plane. Sample or weight for depth-range diversity if your model needs far content.
- Game titles are anonymized as "Game N" in the public metadata; real titles are disclosed under the full-dataset agreement.
- Personal and sensitive information: audio is game audio from the title process — no microphone or player voice is captured — and no personally identifying information about the players ships in any stream or metadata file.
Licensing
All gameplay is recorded under exclusive licenses with the rights holders and captured by consenting, compensated players. Access is gated; request the track you need in the access form:
- Internal Evaluation License: 90-day internal evaluation - train and evaluate models solely to assess the data's value. No publication or release obligation, no deployment or production use. At the end of the period, delete the data and evaluation weights, or convert to a commercial agreement.
- Full dataset / commercial: production training and deployment rights defined per agreement. Contact Origin Lab to license.
No redistribution of the data in any form. See LICENSE.md for the complete terms.
Citation
@misc{originlab2026gameplaycore,
title = {OriginLab Gameplay-Core: Frame-Synced RGB-D Gameplay with
Actions, Camera Pose, and Mechanic Annotations},
author = {Origin Lab},
year = {2026},
url = {https://app.originlab.ai}
}
- Downloads last month
- 124