control-resonant / README.md
XvKuoMing's picture
Dataset card: size category
c48734a verified
|
Raw History Blame Contribute Delete
6.52 kB
metadata
license: other
license_name: gameplay-footage
license_link: LICENSE.md
pretty_name: CONTROL Resonant gameplay with gamepad input
task_categories:
  - video-classification
  - robotics
tags:
  - gameplay
  - gamepad
  - video
  - imitation-learning
  - inverse-dynamics-model
size_categories:
  - n<1K
configs:
  - config_name: clips
    default: true
    data_files:
      - split: train
        path: clips/**

CONTROL Resonant: gameplay video with gamepad input

7.6 hours of one person playing CONTROL Resonant with an Xbox controller, recorded at 1080p and 60 frames per second with game sound. Every video frame comes with the state of the controller at that moment: both sticks, both triggers and 14 buttons.

The data is meant for training models that connect what is on screen to what the player does: inverse dynamics models (video in, input out) and playing agents (video in, next input out). A model trained on it is available as control-resonant-idm.

This is a single-game dataset. Other games get their own dataset in the games_datasets collection.

At a glance

Video 242 clips, 7.62 hours, 1,646,874 frames
Format MP4, H.264, 1920x1080, 60 fps, stereo AAC game sound
Clip length up to 2 minutes
Input Xbox controller, one row per video frame
Sessions 11 play sessions on 5 days
Controller in use 57% of all frames
Size 38 GB

Files

clips/
  metadata.csv        one row per clip
  s01/000.mp4         session 1, clip 0
  s01/001.mp4
  ...
actions/
  s01.parquet         controller state for every frame of every clip in session 1
  ...

Clips of one session follow each other in time. Where something was removed between two clips (see "What was removed"), session_start_s shows the gap.

clips/metadata.csv

column meaning
file_name path of the clip inside clips/
session session id, s01 to s11
clip number of the clip inside its session
recorded date of the session
session_start_s where the clip starts in the original recording, in seconds
duration_s length of the clip in seconds
frames number of video frames
active_fraction share of frames with a button held or a stick or trigger moved
masked_s seconds in which a corner of the picture is blacked out (see below)

actions/sNN.parquet

One row per video frame. Row frame = k of a clip describes the controller while frame k of that clip's video was on screen.

column type meaning
clip string clip id, for example s03/012; matches clips/s03/012.mp4
frame int frame number inside the clip, from 0
time_s float frame / 60
left_x, right_x float stick, -1 (left) to 1 (right)
left_y, right_y float stick, -1 (up) to 1 (down)
lt, rt float trigger, 0 (released) to 1 (fully pressed)
a, b, x, y bool face buttons
lb, rb bool bumpers
left_stick_click, right_stick_click bool stick presses
dpad_up, dpad_down, dpad_left, dpad_right bool D-pad
start, back bool Menu and View buttons

Sticks and triggers have the standard XInput dead zones applied, so a stick at rest reads exactly 0.

Sessions

session recorded clips minutes controller in use
s01 2026-10-28 15 25.8 67%
s02 2026-10-30 11 21.4 62%
s03 2026-11-13 37 70.9 53%
s04 2026-11-13 32 59.0 55%
s05 2026-11-13 49 93.0 42%
s06 2026-11-14 13 24.1 75%
s07 2026-11-14 12 23.3 70%
s08 2026-11-14 29 55.2 64%
s09 2026-11-15 7 13.6 62%
s10 2026-11-15 26 50.2 54%
s11 2026-11-15 11 20.8 66%

Loading

Controller data only (5 MB):

from datasets import load_dataset

actions = load_dataset("parquet", data_files="hf://datasets/XvKuoMing/control-resonant/actions/*.parquet", split="train")
print(actions[0])

One clip with its controller rows:

import pandas as pd
from huggingface_hub import hf_hub_download

repo = "XvKuoMing/control-resonant"
video = hf_hub_download(repo, "clips/s03/012.mp4", repo_type="dataset")
table = pd.read_parquet(hf_hub_download(repo, "actions/s03.parquet", repo_type="dataset"))
rows = table[table.clip == "s03/012"]   # row k belongs to video frame k

Everything:

from huggingface_hub import snapshot_download

snapshot_download("XvKuoMing/control-resonant", repo_type="dataset", local_dir="control-resonant")

How it was recorded

The video is a screen capture of the game at 1080p60. A separate logger read the controller many times per frame and wrote time-stamped changes. The two were aligned on the capture hotkey press, and the controller state was then sampled at the time of each video frame.

What was removed

The recordings were cleaned so that only the game is left:

  • Stretches where the game was not on screen (other windows, the desktop) were cut out, together with a few seconds on each side. Any stretch with keyboard or mouse activity was cut as well. About 12 minutes in total.
  • Message notifications that popped up over the game are covered by a black box in the bottom-right corner, and the sound is muted while the box is on screen. This affects 43 clips for about 6 minutes in total; masked_s gives the amount per clip. The rest of the picture and the controller data in those seconds are untouched.
  • Keyboard and mouse input is not part of the dataset.

Limitations

  • One player, one game, one controller, so play style and button habits are those of one person.
  • Rare inputs are rare here too: D-pad, stick clicks, Start and the right trigger appear in a small share of frames.
  • Menus, cutscenes and loading screens are included. active_fraction helps to filter them.
  • Video and controller are aligned to within about one frame; a small constant offset per session is possible.
  • Masked corners hide a part of the game HUD for a few seconds.

License

The controller data and the file layout are released under CC BY 4.0. The video and sound show CONTROL Resonant, which belongs to its developer and publisher; the footage is shared for research and is not covered by that license. See LICENSE.md.