File size: 2,374 Bytes
f6f5813
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
# RobotPosterior Dataset

Robotic grasp posterior data collected from Isaac Sim physical validation across multiple rounds.

## Contents

- **84 objects** (OakInk + DexYCB)
- **1861 successful grasps** verified by physics simulation (Franka Panda in Isaac Sim)
- Per-object HDF5 files with full grasp pose data

## HDF5 Schema

Each `{obj_id}_robot_gt.hdf5` contains:

```
/
├── attrs:
│   ├── obj_id          (str)
│   ├── n_successful    (int)
│   └── sources         (str)  e.g. "R1(×5) | R2(×13) | NEW(×20)"

└── successful_grasps/
    ├── attrs: count
    └── grasp_N/
        ├── grasp_point         (3,)   float32  — contact midpoint, canonical mesh frame
        ├── rotation            (3,3)  float32  — gripper rotation matrix
        ├── approach_dir        (3,)   float32  — rotation[:, 2]
        ├── finger_dir          (3,)   float32  — rotation[:, 0]
        ├── [contact_points_local] (2,3) float32 — actual finger tip positions
        └── attrs:
            ├── gripper_width   (float)  meters
            ├── score           (float)
            └── approach_type   (str)
```

## Coordinate Frame

- `grasp_point`, `rotation`: **canonical mesh local frame** (Z-up, object bottom at Z=0)
- `contact_points_local`: world offset from `OBJECT_POSITION` in Isaac Sim

## Data Sources

| Label | Round | Server | Objects | Grasps |
|-------|-------|--------|---------|--------|
| R1 | Round 1 | Titan | 48 | ~231 |
| R2 | Round 2 | Titan | 55 | ~592 |
| R3 | Round 3 | Titan | 43 | ~181 |
| D1 | DexYCB R1 | Titan | 7 | ~14 |
| D2 | DexYCB R2 | Titan | 6 | ~41 |
| NEW | Round 0+1 | RTX5090 | 70 | ~802 |

## Candidate Generation

Grasp candidates generated with **50% Human-Prior guided + 50% random** sampling (v5.1 Raycast scorer).  
All entries physically verified in **Isaac Sim** with Franka Panda robot.

## Usage

```python
import h5py, numpy as np

with h5py.File('C28001_robot_gt.hdf5', 'r') as f:
    print(f"n_successful: {f.attrs['n_successful']}")
    for key in f['successful_grasps'].keys():
        g = f['successful_grasps'][key]
        grasp_point  = g['grasp_point'][:]    # (3,) meters, mesh local frame
        approach_dir = g['approach_dir'][:]   # (3,) unit vector
        width        = g.attrs['gripper_width']
```