ASA / README.md
Joringell's picture
Upload README.md
fb4b281 verified
# Dataset for ASA
This repo provides the data used in our paper ***Act, Sense, Act: Learning Non-Markovian Active Perception Strategies from Large-Scale Egocentric Human Data***. It consists of a curated combination of public egocentric human datasets and collected robot data, processed into a unified format for training.
For more details, please refer to the [paper](https://arxiv.org/abs/2602.04600) and [project page](https://jern-li.github.io/asa/).
## Dataset Overview
### Human Data
| Source | Type | Samples | Takes | Languages | Take_Languages
|-----------------|-----------------------------------|----------------|----------------|----------------|----------------
|CaptainCook4D | frame3_chunk1-100_his10-15_anno_image | 1,071,604 | 257 | 351 | 3417
|EgoExo4D | proprio_frame1_chunk3-100_his30-15_image | 421,582 | 249 | 2730 | 3131
### Robot Data
| Source | Task | Type | Samples | Takes | Languages | Take_Languages | Details |
|-----------------|-----------|------------------------|----------------|----------------|----------------|----------------|------------------------
|Monte02 | task1_1 | frame1_chunk3-100_his30-15_extend90_gripper_image | 493,678 | 191 | 3 | 573 | vision/current_image 224, vision/history_image 224, hf.feature
|Monte02 | task3_1 | frame1_chunk3-100_his30-15_extend90_gripper_image_new | 383,729 | 102 | 3 | 306 | new (no light), new_anno
|Monte02 | task3_2 |frame1_chunk3-100_his30-15_extend90_gripper_image_newnew_aug | 300,958 | 83 | 3 | 249 | new sub23, new sub1 + new-only-sub1(18w), and img-aug
|Monte02 | task1_2 | frame1_chunk3-100_his30-15_extend90_gripper_image_move | 375,143 | 188 | 2 | 376 | only subtask 2 and 3, and source = 'Monte02_Move'
|Monte02 | task1_2 | frame1_chunk3-100_his30-15_extend90_gripper_hand_new |275, 699 | 218 | 2 | 218 | sub1 old + sub4 new, source='Monte02', 'Monte02_12sub4'
|Monte02 | task2_1 | proprio_frame1_chunk3-100_his30-15_extend90_gripper_newdata_image_new | 151,628 | 69 | 2 | 138 | new data, big ring
## Dataset Structure
### Directory Layout
```
ASA/
├── captaincook4d
│   └── hf_datasets
│   └── proprio_frame3_chunk1-100_his30-15_anno_image
│   ├── by_language.pkl
│   ├── by_take_language.pkl
│   ├── by_take.pkl
│   ├── data-00000-of-00028.arrow
│   ├── ....
│   ├── data-00027-of-00028.arrow
│   ├── dataset_info.json
│   └── state.json
├── egoexo4d
│   └── hf_datasets
│   └── proprio_frame1_chunk3-100_his30-15_image
│   ├── xxx.pkl
│   ├── xxxxx.arrow
│   └── xxx.json
└── monte02
   ├── hf_datasets
   │   ├── task1_1
   │   │   └── proprio_xxx
   │   │   ├── xxx.pkl
   │   │   ├── xxxxx.arrow
   │   │   └── xxx.json
   │   ├── task1_2
   │   ├── task2_1
   │   ├── task3_1
   │   └── task3_2
   └── raw_data
   ├── task1_1.zip
   │   └── folder
   │   └── sample_xxxx_xxx
   │   ├── annotation.json
   │   ├── head_video.avi
   │   ├── robot_data.h5
   │   ├── label_result.txt (optional, not available for all samples)
   │   ├── left_video.avi (optional)
   │   ├── right_video.avi (optional)
  │   └── valid.txt
   ├── task1_2.zip
   ├── task2_1.zip
  ├── task3_1.zip
   └── task3_2.zip
```
### Data Fields
<details>
<summary> CaptainCook4D </summary>
| Key | Type | Shape | Details |
|--------------------------------------|----------------|----------------|--------------------------------------
| `source` | `str` | - | from which dataset
| `take_name` | `str` | - |
| `frame_idx` | `int` | - | index of the frame in the filtered take (not continuous) (aligned with pose index)
| `vision/rgb_image` | `bytes` | - |RGB image of size **(504, 896, 3)**
`vision/current_image` |`Image` (hf.feature) | - |head RGB image of size **(224, 224, 3)**
`vision/history_image` |`list(Image)` (hf.feature) | - | 5 history (5s, t-5 ~ t-1) head RGB image of size **(224, 224, 3)**
| `vision/video_frame` | `int` | - |index of the frame in the video
| `vision/histroy_idx` | `list` | - |index of the histroy in the **HF_IMAGE_DATASET** , maybe in past subtask
`current/complete ` | `bool` | - | whether the subtask is complete
| `annotation/language` | `str` | - |
| `annotation/start_frame` | `int` | - |start_frame of this keystep
| `annotation/end_frame` | `int` | - |
| `annotation/delta_idx` | `int` | - | index change in the filtered keystep
| `current/head/raw_pose` | `ndarray` | (4, 4) | in the world frame
| `current/left_hand/raw_pose` | `ndarray` | (26, 4, 4) | 26 joints of the left hand
| `current/left_hand/mano_params` | `ndarray` | (15,) | not use
| `current/right_hand/raw_pose` | `ndarray` | (26, 4, 4) |
| `current/right_hand/mano_params` | `ndarray` | (15,) |
| `current/head/pose_in_base` | `ndarray` | (9,) | in the base frame
| `current/left_hand/pose_in_base` | `ndarray` | (26, 9) | all 26 joints
| `current/left_hand/wrist_in_base` | `ndarray` | (9,) | only wrist
| `current/left_hand/gripper` | `ndarray` | (1,) |
| `current/right_hand/pose_in_base` | `ndarray` | (26, 9) | all 26 joints
| `current/right_hand/wrist_in_base` | `ndarray` | (9,) |
| `current/right_hand/gripper` | `ndarray` | (1,) | normalized gripper state
| `current/head/move ` | `bool` | - | whether the component is moving in current subtask
| `current/left_hand/move` | `bool` | - |
| `current/right_hand/move` | `bool` | - |
| `history/complete` | `ndarray` | (100,) | future chunk 100
| `history/head/move` | `ndarray` | (100,) |
| `future/head/pose_in_base` | `ndarray` | (100, 9) |
| `future/left_hand/move ` | `ndarray` | (100,) |
| `future/left_hand/wrist_in_base` | `ndarray` | (100, 9) |
| `future/left_hand/gripper` | `ndarray` | (100,1) |
| `future/right_hand/move ` | `ndarray` | (100,) |
| `future/right_hand/wrist_in_base` | `ndarray` | (100, 9) |
| `future/right_hand/gripper` | `ndarray` | (100,1) |
| `history/complete` | `list` | - | history chunk 15, only in this subtask
| `history/head/move` | `list` | - |
| `history/head/pose_in_base` | `list` | - |
| `history/left_hand/move ` | `list` | - |
| `history/left_hand/wrist_in_base` | `list` | - |
| `history/left_hand/gripper` | `list` | - |
| `history/right_hand/move ` | `list` | -
| `history/right_hand/wrist_in_base` | `list` | - |
| `history/right_hand/gripper` | `list` | - |
</details>
<details>
<summary> EgoExo4D </summary>
| Key | Type | Shape | Details |
|--------------------------------------|----------------|----------------|--------------------------------------
| `source` | `str` | - | from which dataset
| `take_name` | `str` | - | |
| `frame_idx` | `int` | - | index of the frame in the filtered take (not continuous) |
| `vision/rgb_image` | `bytes` | - |RGB image of size **(1408, 1408, 3)**
`vision/current_image` |`Image` (hf.feature) | - |head RGB image of size **(224, 224, 3)**
`vision/history_image` |`list(Image)` (hf.feature) | - | 5 history (5s, t-5 ~ t-1) head RGB image of size **(224, 224, 3)**
| `vision/video_frame` | `int` | - |index of the frame in the video
| `vision/histroy_idx` | `list` | - |index of the histroy in the **HF_IMAGE_DATASET** |
| `annotation/language` | `str` | - | coarse_grained or fine_grained |
| `annotation/start_frame` | `int` | - |start_frame of this keystep |
| `annotation/end_frame` | `int` | - |
| `annotation/delta_idx` | `int` | - | index change in the filtered keystep |
| `current/head/raw_pose` | `ndarray` | (4, 4) | in the world frame |
| `current/left_hand/raw_position` | `ndarray` | (26, 3) | 26 joints of the left hand
| `current/left_hand/mano_params` | `ndarray` | (15,) |
| `current/left_hand/wrist_pose` | `ndarray` | (4,4) |wrist pose of left hand, rotation is optimized by MANO
| `current/right_hand/raw_position` | `ndarray` | (26, 3) |
| `current/right_hand/mano_params` | `ndarray` | (15,) |
| `current/right_hand/wrist_pose` | `ndarray` | (4,4) |
| `current/head/pose_in_base` | `ndarray` | (9,) | in the base frame|
| `current/left_hand/wrist_in_base` | `ndarray` | (9,) | only wrist
| `current/left_hand/gripper` | `ndarray` | (1,) | gripper width
| `current/right_hand/wrist_in_base` | `ndarray` | (9,) |
| `current/right_hand/gripper` | `ndarray` | (1,) |
| `current/head/move ` | `bool` | - | whether the component is moving in current subtask
| `current/left_hand/move` | `bool` | - |
| `current/right_hand/move` | `bool` | - |
| `history/complete` | `ndarray` | (100,) | future chunk 100
| `history/head/move` | `ndarray` | (100,) |
| `future/head/pose_in_base` | `ndarray` | (100, 9) |
| `future/left_hand/move ` | `ndarray` | (100,) |
| `future/left_hand/wrist_in_base` | `ndarray` | (100, 9) |
| `future/left_hand/gripper` | `ndarray` | (100,1) |
| `future/right_hand/move ` | `ndarray` | (100,) |
| `future/right_hand/wrist_in_base` | `ndarray` | (100, 9) |
| `future/right_hand/gripper` | `ndarray` | (100,1) |
| `history/complete` | `list` | - | history chunk 15
| `history/head/move` | `list` | - |
| `history/head/pose_in_base` | `list` | - |
| `history/left_hand/move ` | `list` | - |
| `history/left_hand/wrist_in_base` | `list` | - |
| `history/left_hand/gripper` | `list` | - |
| `history/right_hand/move ` | `list` |-
| `history/right_hand/wrist_in_base` | `list` | - |
| `history/right_hand/gripper` | `list` | - |
</details>
<details>
<summary> Monte02 </summary>
| Key | Type | Shape | Details
|--------------------------------------|----------------|----------------|-------------------------
`source` | `str` | - |
`take_name` | `str` | - | sample_...
`frame_idx` |`int` | - |
`vision/video_frame` | `int` | - |
`vision/rgb_image` | `bytes` | - |head RGB image of size **(640, 480, 3)**
`vision/current_image` |`Image` (hf.feature) | - |head RGB image of size **(224, 224, 3)**
`vision/history_image` |`list(Image)` (hf.feature) | - | 5 history (5s, t-5 ~ t-1) head RGB image of size **(224, 224, 3)**
`vision/history_idx ` | `list` | - | [t-15 ~ t]
`annotation/task` | `str` | - | task language
`annotation/language` | `str` | - | subtask language
`annotation/start_frame` | `int` | - |
`annotation/end_frame` | `int` | - |
`annotation/delta_idx` | `int` | - |
`current/complete ` | `bool` | - | whether the subtask is complete
`current/left_hand/gripper ` | `ndarray` | (1,) | 0 or 1 (? 0.065)
`current/right_hand/gripper` | `ndarray` | (1,) | 0 or 1 (? 0.065)
`current/left_hand/gripper_width ` | `ndarray` | (1,) | 0~0.01
`current/right_hand/gripper_width` | `ndarray` | (1,) | 0~0.01
`current/head/angles` | `ndarray` | (2,) | pitch, yaw
`current/chassis/pose_in_init` | `ndarray` | (7,) | xyz + wxyz
`current/head/pose_in_base` | `ndarray` | (9,) | xyz + rot6d, base = init_head
`current/head/pose_in_step_base` | `ndarray` | (9,) | xyz + rot6d, step_base = current init_head
`current/left_hand/wrist_in_base` | `ndarray` | (9,)
`current/right_hand/wrist_in_base ` | `ndarray` | (9,)
`current/left_hand/wrist_in_step_base` | `ndarray` | (9,)
`current/right_hand/wrist_in_step_base` | `ndarray` | (9,)
`current/head/move ` | `bool` | - | whether the component is moving in current subtask
`current/left_hand/move` | `bool` | - |
`current/right_hand/move` | `bool` | - |
`future/complete` | `ndarray` | (100,) |future actions and states
`future/head/move` | `ndarray` | (100,) |
`future/head/pose_in_base` | `ndarray` | (100, 9)
`future/head/pose_in_step_base ` | `ndarray` | (100, 9)
`future/left_hand/move` | `ndarray` | (100,)
`future/left_hand/wrist_in_base` | `ndarray` | (100, 9)
`future/left_hand/wrist_in_step_base` | `ndarray` | (100, 9)
`future/left_hand/gripper ` | `ndarray` | (100, 1)
`future/right_hand/move ` | `ndarray` | (100,)
`future/right_hand/wrist_in_base ` | `ndarray` | (100, 9)
`future/right_hand/wrist_in_step_base ` | `ndarray` | (100, 9)
`future/right_hand/gripper` | `ndarray` | (100, 1)
`history/complete` | `list` | - | history actions and states
`history/head/move` | `list` | - |
`history/head/pose_in_base` | `list` | - |
`history/head/pose_in_step_base ` | `list` | -|
`history/left_hand/move` | `list` | -|
`history/left_hand/wrist_in_base` | `list` | -|
`history/left_hand/wrist_in_step_base` | `list` | -|
`history/left_hand/gripper ` | `list` | - |
`history/right_hand/move ` | `list` | - |
`history/right_hand/wrist_in_base ` | `list` | -|
`history/right_hand/wrist_in_step_base ` | `list` | -|
`history/right_hand/gripper` | `list` | -|
</details>
## Notes
- We provide preprocessed datasets to ensure consistent quality and reduce preprocessing overhead.
- Human data is filtered with strict criteria to improve learning stability.
- Robot data is collected in real-world environments.