cyberorigin/cyber_take_the_item · Datasets at Hugging Face

token_dtype stringclasses 1 value	s int64 16 16	h int64 16 16	w int64 16 16	vocab_size int64 262k 262k	hz int64 30 30	tokenizer_ckpt stringclasses 1 value	num_images int64 218k 399k
uint32	16	16	16	262,144	30	imagenet_256_L.ckpt	282,241
uint32	16	16	16	262,144	30	imagenet_256_L.ckpt	312,472
uint32	16	16	16	262,144	30	imagenet_256_L.ckpt	285,401
uint32	16	16	16	262,144	30	imagenet_256_L.ckpt	233,656
uint32	16	16	16	262,144	30	imagenet_256_L.ckpt	217,965
uint32	16	16	16	262,144	30	imagenet_256_L.ckpt	398,778
uint32	16	16	16	262,144	30	imagenet_256_L.ckpt	380,849
uint32	16	16	16	262,144	30	imagenet_256_L.ckpt	368,323
uint32	16	16	16	262,144	30	imagenet_256_L.ckpt	352,934
uint32	16	16	16	262,144	30	imagenet_256_L.ckpt	246,178
uint32	16	16	16	262,144	30	imagenet_256_L.ckpt	250,753
uint32	16	16	16	262,144	30	imagenet_256_L.ckpt	288,209

token_dtype

stringclasses

1 value

int64

vocab_size

int64

262k

int64

tokenizer_ckpt

stringclasses

1 value

num_images

int64

218k

399k

uint32

262,144

imagenet_256_L.ckpt

282,241

uint32

262,144

imagenet_256_L.ckpt

312,472

uint32

262,144

imagenet_256_L.ckpt

285,401

uint32

262,144

imagenet_256_L.ckpt

233,656

uint32

262,144

imagenet_256_L.ckpt

217,965

uint32

262,144

imagenet_256_L.ckpt

398,778

uint32

262,144

imagenet_256_L.ckpt

380,849

uint32

262,144

imagenet_256_L.ckpt

368,323

uint32

262,144

imagenet_256_L.ckpt

352,934

uint32

262,144

imagenet_256_L.ckpt

246,178

uint32

262,144

imagenet_256_L.ckpt

250,753

uint32

262,144

imagenet_256_L.ckpt

288,209

CyberOrigin Dataset

Our data includes information from home services, the logistics industry, and laboratory scenarios. For more details, please refer to our Offical Data Website

contents of the dataset:

cyber_take_the_item # dataset root path
  └── data/
      ├── metadata_ID1_240808.json
      ├── segment_ids_ID1_240808.bin # for each frame segment_ids 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).
      ├── videos_ID1_240808.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 magvit2.ckpt weights.
      ├── ...
  └── ...

{
    "task": "Take the Item",
    "total_episodes": 18779,
    "total_frames": 3617759,
    "token_dtype": "uint32",
    "vocab_size": 262144,
    "fps": 30,
    "manipulation_type": "Single Arm",
    "language_annotation": "None",
    "scene_type": "Table Top",
    "data_collect_method": "Directly Collection on Human"
}