1x-technologies/worldmodel

Dataset for the 1X World Model Challenge.

Download with:

huggingface-cli download 1x-technologies/worldmodel --repo-type dataset --local-dir data

Changes from v1.1:

• New train and val dataset of 100 hours, replacing the v1.1 datasets
• Blur applied to faces
• Shared a new raw video dataset under CC-BY-NC-SA 4.0: https://huggingface.co/datasets/1x-technologies/worldmodel_raw_data
• Example scripts to decode Cosmos Tokenized bins cosmos_video_decoder.py and load in frame data unpack_data.py

Contents of train/val_v2.0:

The training dataset is shareded into 100 independent shards. The definitions are as follows:

• video_{shard}.bin: 8x8x8 image patches at 30hz, with 17 frame temporal window, encoded using NVIDIA Cosmos Tokenizer "Cosmos-Tokenizer-DV8x8x8".

• segment_idx_{shard}.bin - Maps each frame i to its corresponding segment index. You may want to use this to separate non-contiguous frames from different videos (transitions).

• states_{shard}.bin - States arrays (defined below in Index-to-State Mapping) stored in np.float32 format. For frame i, the corresponding state is represented by states_{shard}[i].

• metadata - The metadata.json file provides high-level information about the entire dataset, while metadata_{shard}.json files contain specific details for each shard.

  Index-to-State Mapping (NEW)

   {
        0: HIP_YAW
        1: HIP_ROLL
        2: HIP_PITCH
        3: KNEE_PITCH
        4: ANKLE_ROLL
        5: ANKLE_PITCH
        6: LEFT_SHOULDER_PITCH
        7: LEFT_SHOULDER_ROLL
        8: LEFT_SHOULDER_YAW
        9: LEFT_ELBOW_PITCH
        10: LEFT_ELBOW_YAW
        11: LEFT_WRIST_PITCH
        12: LEFT_WRIST_ROLL
        13: RIGHT_SHOULDER_PITCH
        14: RIGHT_SHOULDER_ROLL
        15: RIGHT_SHOULDER_YAW
        16: RIGHT_ELBOW_PITCH
        17: RIGHT_ELBOW_YAW
        18: RIGHT_WRIST_PITCH
        19: RIGHT_WRIST_ROLL
        20: NECK_PITCH
        21: Left hand closure state (0 = open, 1 = closed)
        22: Right hand closure state (0 = open, 1 = closed)
        23: Linear Velocity
        24: Angular Velocity
    }
  

Previous version: v1.1

• magvit2.ckpt - weights for MAGVIT2 image tokenizer we used. We provide the encoder (tokenizer) and decoder (de-tokenizer) weights.

Contents of train/val_v1.1:

• video.bin - 16x16 image patches at 30hz, each patch is vector-quantized into 2^18 possible integer values. These can be decoded into 256x256 RGB images using the provided magvig2.ckpt weights.
• segment_ids.bin - for each frame segment_ids[i] uniquely points to the segment index that frame i came from. You may want to use this to separate non-contiguous frames from different videos (transitions).
• actions/ - a folder of action arrays stored in np.float32 format. For frame i, the corresponding action is given by joint_pos[i], driving_command[i], neck_desired[i], and so on. The shapes and definitions of the arrays are as follows (N is the number of frames):
  • joint_pos (N, 21): Joint positions. See Index-to-Joint Mapping below.
  • driving_command (N, 2): Linear and angular velocities.
  • neck_desired (N, 1): Desired neck pitch.
  • l_hand_closure (N, 1): Left hand closure state (0 = open, 1 = closed).
  • r_hand_closure (N, 1): Right hand closure state (0 = open, 1 = closed).

  Index-to-Joint Mapping (OLD)

   {
        0: HIP_YAW
        1: HIP_ROLL
        2: HIP_PITCH
        3: KNEE_PITCH
        4: ANKLE_ROLL
        5: ANKLE_PITCH
        6: LEFT_SHOULDER_PITCH
        7: LEFT_SHOULDER_ROLL
        8: LEFT_SHOULDER_YAW
        9: LEFT_ELBOW_PITCH
        10: LEFT_ELBOW_YAW
        11: LEFT_WRIST_PITCH
        12: LEFT_WRIST_ROLL
        13: RIGHT_SHOULDER_PITCH
        14: RIGHT_SHOULDER_ROLL
        15: RIGHT_SHOULDER_YAW
        16: RIGHT_ELBOW_PITCH
        17: RIGHT_ELBOW_YAW
        18: RIGHT_WRIST_PITCH
        19: RIGHT_WRIST_ROLL
        20: NECK_PITCH
    }
  

We also provide a small val_v1.1 data split containing held-out examples not seen in the training set, in case you want to try evaluating your model on held-out frames.