cyberorigin/cyber_fold_towels_v0.1 · 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 7.64k 504k	num_episodes int64 30 1.22k	task stringclasses 1 value
uint32	16	16	16	262,144	30	data/magvit2.ckpt	451,010	761	fold_towels
uint32	16	16	16	262,144	30	data/magvit2.ckpt	142,409	228	fold_towels
uint32	16	16	16	262,144	30	data/magvit2.ckpt	504,486	1,039	fold_towels
uint32	16	16	16	262,144	30	data/magvit2.ckpt	502,831	1,224	fold_towels
uint32	16	16	16	262,144	30	data/magvit2.ckpt	462,158	1,181	fold_towels
uint32	16	16	16	262,144	30	data/magvit2.ckpt	406,734	1,214	fold_towels
uint32	16	16	16	262,144	30	data/magvit2.ckpt	7,638	30	fold_towels
uint32	16	16	16	262,144	30	data/magvit2.ckpt	171,681	480	fold_towels
uint32	16	16	16	262,144	30	data/magvit2.ckpt	302,218	770	fold_towels

token_dtype

stringclasses

1 value

int64

vocab_size

int64

262k

int64

tokenizer_ckpt

stringclasses

1 value

num_images

int64

7.64k

504k

num_episodes

int64

1.22k

task

stringclasses

1 value

uint32

262,144

data/magvit2.ckpt

451,010

761

fold_towels

uint32

262,144

data/magvit2.ckpt

142,409

228

fold_towels

uint32

262,144

data/magvit2.ckpt

504,486

1,039

fold_towels

uint32

262,144

data/magvit2.ckpt

502,831

1,224

fold_towels

uint32

262,144

data/magvit2.ckpt

462,158

1,181

fold_towels

uint32

262,144

data/magvit2.ckpt

406,734

1,214

fold_towels

uint32

262,144

data/magvit2.ckpt

7,638

fold_towels

uint32

262,144

data/magvit2.ckpt

171,681

480

fold_towels

uint32

262,144

data/magvit2.ckpt

302,218

770

fold_towels

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 dataset:

cyber_fold_towels # 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": "Fold Towels",
    "total_episodes": 6927,
    "total_frames": 9938323,
    "token_dtype": "uint32",
    "vocab_size": 262144,
    "fps": 30,
    "manipulation_type": "Bi-Manual",
    "language_annotation": "None",
    "scene_type": "Table Top",
    "data_collect_method": "Directly Collection on Human"
}