Gameplay data for world models and frontier AI.

Parkour and traversal across … shared frame positions. RGB, depth and surface normals are paired with camera pose, the values behind movement, look and jump controls, and recorded character-state changes. Inputs and engine events let you inspect what happened alongside each observation.

Project and map

unavailable

Engine

unavailable

Recorded

2026-10-16T16:57:17.852Z

Capture and preparation tooling

unavailable

Rights

Not supplied for this sample. The displayed package grants no training or redistribution rights.

Loading manifest provenance.

1. 1. Download the telemetry sample

   Original converter-produced tables supplied with the capture. The ZIP contains the unchanged source manifest, original converter QA, the preparation report, notes and three Parquet tables.

   Download telemetry-sample.zip (≈518 KiB)
   • Session manifest · JSON · supplied capture manifest; streams, clock and documented conventions
   • QA report · JSON · preparation validation, not a converter certificate
   • Preparation report · JSON · preview preparation and table provenance, separate from source QA

   The lossless depth and normals files and the full-resolution RGB aren't hosted here. Ask if you need them for evaluation.

2. 2. Unzip it

   manifest.json · unchanged source manifest · qa.json · original converter report · preparation.json · web preparation report · PREPARATION.txt · preparation notes · tables/frames.parquet · original converter table · tables/input_raw.parquet · original converter table · tables/events.parquet · original converter table

3. 3. Read the first rows

   With pyarrow and pandas installed, this reads the frames table.

   first_rows.py Copy

   import pyarrow.parquet as pq
   
   frames = pq.read_table("tables/frames.parquet").to_pandas()
   print(len(frames), "frames")                     # …
   print(frames[["frame_idx", "t_ns", "cam_pos_x", "cam_pos_y", "cam_pos_z"]].head())

   Full example: read all three tables and join them on frame

   quickstart.py Copy

   import pyarrow.parquet as pq
   
   frames = pq.read_table("tables/frames.parquet").to_pandas()
   inputs = pq.read_table("tables/input_raw.parquet").to_pandas()
   events = pq.read_table("tables/events.parquet").to_pandas()
   
   print(len(frames), "frames")                     # …
   print(frames[["frame_idx", "t_ns", "cam_pos_x", "cam_pos_y", "cam_pos_z"]].head())
   
   # Join observations by frame index, never elapsed time.
   # Preserve source boundary events N; these have no matching frame row.
   aligned_events = events.merge(frames[["frame_idx", "t_ns"]],
       on="frame_idx", how="left", suffixes=("_event", "_frame"))
   aligned_inputs = inputs.merge(frames[["frame_idx"]],
       left_on="frame_next", right_on="frame_idx", how="left")
   print(aligned_events.head(), aligned_inputs.head(), sep="\n")

Frame alignment and playback time

Every stream shares the same frame index. This preview plays at 60 FPS; the original capture timestamps preserve the game's timing. Use frame numbers to match video with telemetry.

Media time

The preview video's clock: … frames at a constant 60 FPS, so … seconds.

Capture time

The game's own clock, which ran for … seconds because real frames weren't evenly spaced.

From a media timestamp:

N = floor(mediaTime × 60)

Don't seek video by capture time, and don't scale by the duration ratio.

Decoding and conventions

Use the selected manifest's documented stream encodings, decode formulas and coordinate conventions. Unspecified conventions are unavailable.

The videos on this page are downscaled, lossy previews. The depth preview shows the log codes as greyscale, so it's a picture of depth rather than depth values. Use the lossless source files for anything numerical.

Real records

Available records exactly as they appear in the selected sample. Use the arrow keys to move between tabs.

Source capture checks

Loading the source converter report…

Preview preparation

Loading preview preparation checks… Separate preparation checks verify preview frame counts and source mapping. The original manifest, converter QA and Parquet tables remain unchanged. At missing RGB positions, the preview holds the previous recorded image; availability flags and hold provenance retain the distinction.

How it was captured

An engine plugin records viewport images, camera pose, player inputs, mapped actions and available state and events on a shared frame index. Depth and surface normals are captured from the engine's rendering buffers. The supplied manifest documents the stream encodings, camera conventions and coordinate system.

Sample scope

• One Parkour Race evaluation capture.
• No audio or post-HUD RGB in this sample.
• Downscaled, lossy web previews; use original visual files for numerical work.
• Missing RGB captures are explicit in source availability and preview hold provenance.
• Not supplied for this sample. The displayed package grants no training or redistribution rights.