szcharlesji/airhockey-1m

Air Hockey SA v6 (1M)

Action-conditioned, 50 Hz air-hockey trajectories for training world models, video prediction systems, visual dynamics models, and object-centric models. The dataset contains one million 128 x 128 RGB frames from a two-player MuJoCo air-hockey simulator, with commanded planar mallet actions for both players.

At a glance

Property	Value
Episodes	1,000 fixed-length episodes
Frames	1,000,000 total (1,000 per episode)
Frame rate	50 Hz
Image format	RGB uint8, shape (T, 128, 128, 3)
Action format	float32, shape (T, 4)
Action layout	[p1_x, p1_y, p2_x, p2_y] commanded mallet position in world coordinates
Train / validation	950 / 50 whole episodes (950,000 / 50,000 frames)
Storage	Manifest-backed HDF5 shards, langtable_action_h5_v1

The validation split is held out by complete game, not by frame. It contains 25 aggressive and 25 smooth-random episodes.

Data generation

The data was collected with the Air Hockey Challenge 2023 tournament environment. Each source contributes exactly half of the episodes:

Source	Episodes	Players
aggressive	500	tournament_aggressive vs. itself
random	500	smooth_random vs. itself

All episodes are exactly 1,000 simulator/control frames (20 seconds at 50 Hz). The collector does not terminate an episode on scoring, centre-stuck, or edge events. A scored puck remains at its goal coordinate, has contacts disabled, and is hidden in subsequent rendered frames; both players then use fresh smooth_random policies for the remaining tail. Other fault, stuck, and edge events remain visible and keep simulating. This avoids a length/termination shortcut for predictive models.

Physical and visual parameters

• Puck radius: 0.10 m.
• Physical mallet radius: 0.10 m.
• Robot arm visual scale: 1.5x (visual meshes only; action space and collision geometry are otherwise unchanged).
• Puck orientation marker: a fixed-size asymmetric red cross.
• Goal mouth is scaled from puck size with a constant clearance.
• The collector uses a hard mallet level/height safety projection to prevent mallets from tilting into the table.

Both policies have independent intermittent holds. On each active 50 Hz control step, a hold starts with probability 0.005; a hold lasts uniformly from 50 to 250 frames (1 to 5 seconds). During a hold, the last safe policy-generated target is retained with zero desired velocity and the policy state is frozen.

HDF5 layout

h5_manifest.json is authoritative; do not glob shard files when loading. It maps each episode key to a split, shard, and length. Shards contain:

train/shard_*.h5
valid/shard_*.h5

/episodes/<episode_key>/images   uint8   (T, 128, 128, 3)
/episodes/<episode_key>/actions  float32 (T, 4)

Images are stored as RGB values in [0, 255]; normalize to [0, 1] or your model's preferred range at load time. Actions are commanded mallet XY targets in simulator world coordinates, not pixels, joint angles, velocities, or forces. They are aligned to the transition convention: actions[t] is the command that drives the world from image t toward image t + 1.

The provided SA loader reads this layout directly with dataset: languagetable. Other frameworks can load it with h5py using the manifest as the index.

import json
from pathlib import Path
import h5py

root = Path("airhockey_sa_h5_puck100_mallet100_arm15_v6_1m")
manifest = json.loads((root / "h5_manifest.json").read_text())
entry = manifest["splits"]["train"][0]

with h5py.File(root / entry["shard"], "r") as f:
    episode = f["episodes"][entry["key"]]
    images = episode["images"][:]   # uint8, (1000, 128, 128, 3)
    actions = episode["actions"][:] # float32, (1000, 4)

Recommended uses

• Action-conditioned next-frame, video, or latent-state prediction.
• Object-centric perception and slot-based dynamics.
• Latent world models, planning, and model-based RL pretraining.
• Video representation learning with controlled interventions.
• Long-horizon rollout evaluation: each episode provides 20 seconds of 50 Hz dynamics rather than terminating at a score.

For models requiring fixed clips, sample contiguous windows within an episode. For example, SA uses 20-frame dense windows. Keep temporal order intact and split by episode; do not randomly split frames across train and validation.

Caveats

• This is simulated data, not real robot or camera data. Transfer to physical air hockey will require domain adaptation and safety validation.
• The background/table and robot configuration are intentionally consistent; appearance diversity is limited relative to real-world video.
• The goal-tail rule deliberately makes the puck visually absent after a real score while the episode continues. Models should treat this as an observed part of the environment, not a missing/corrupted frame.
• Long idle holds are intentional action-conditioned behavior, not duplicate data or dropped frames.
• No license has been declared in this directory. Set the intended Hugging Face license and attribution terms before publishing.

Provenance and integrity

The source YAML is configs/data/airhockey-v6.yaml in the associated collection repository. The exact collection content hash is 8ccef8cb08a33fc12bfdfd15064c39c2f7c8c1e3f727a6d9337176b0356ffde7. collection_config.json records the completed collection and conversion metadata. The manifest is marked complete: true; it records 1,000,000 frames and zero skipped short episodes.