File size: 5,057 Bytes
d1a914c
f74e0c6
 
 
 
 
 
 
 
 
 
 
d1a914c
f74e0c6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- robotics
- isaac-lab
- forge
- manipulation
- peg-insertion
- force-torque
---

# Forge v1 — contact-rich manipulation with force/torque, 3 tasks

Expert demonstrations for the three [Isaac Lab](https://github.com/isaac-sim/IsaacLab) **FORGE**
tasks, collected from trained `rl_games` PPO policies. Successful episodes only; the failures from
the same collection are published separately as
[`angledusgar/forge-failure-v1`](https://huggingface.co/datasets/angledusgar/forge-failure-v1).

Beyond the usual LIBERO-shaped fields this dataset carries the **wrist force/torque signal**, which
is what makes contact-rich insertion legible: the peg/hole clearance is 0.114 mm, far below what a
224×224 camera can resolve, so the moment of contact and the onset of a jam are visible in the
wrench long before they are visible in pixels.

## Contents

| task | episodes | steps/episode | prompt |
|---|---|---|---|
| `Isaac-Forge-PegInsert-Direct-v0` | ~984 | 149 | "insert the peg into the socket" |
| `Isaac-Forge-GearMesh-Direct-v0` | ~996 | 299 | "mesh the gear onto the shaft" |
| `Isaac-Forge-NutThread-Direct-v0` | ~988 | 449 | "thread the nut onto the bolt" |
| **total** | **2,968** | — | 888,032 frames |

All three tasks live in one repository, distinguished by the per-frame task string
(`meta/tasks.jsonl`), the same packing `physical-intelligence/libero` uses.

## Features

| key | dtype | shape | meaning |
|---|---|---|---|
| `image` | video (AV1) | 256×256×3 | fixed third-person camera, one per environment |
| `wrist_image` | video (AV1) | 256×256×3 | wrist camera, mounted on `panda_hand` |
| `state` | float32 | (8,) | `eef_pos` (3) + axis-angle from `eef_quat` (3) + gripper joints (2) |
| `actions` | float32 | (7,) | the executed Forge environment action |
| **`ft`** | float32 | (6,) | wrench at the wrist in the EE frame — force (3) then torque (3), smoothed |
| **`ft_noisy`** | float32 | (3,) | the force the policy actually observed, with Forge's observation noise |

`state` is deliberately 8-D and named as in the openpi LIBERO conversion, so
`LeRobotLiberoDataConfig` can be reused unchanged; `ft` / `ft_noisy` are extra keys that openpi's
`RepackTransform` simply drops. Robot is a Franka Panda; fps 15 (Forge control rate); LeRobot
dataset format `v2.1`.

Two force channels are kept on purpose: a detector can be *trained* on the clean wrench while being
*evaluated* on the noisy signal the robot really had.

## How it was collected

Sixteen environments were stepped in lockstep — Factory/Forge resets all environments together, so
each window yields exactly N equal-length episodes — with the expert acting deterministically.
Episodes were kept regardless of outcome and split by result afterwards.

| task | expert checkpoint | success rate |
|---|---|---|
| PegInsert | 300-epoch PPO, trained 2026-06-30 on the **stock** task | ~96% |
| GearMesh | 300-epoch PPO, trained 2026-08-01 | ~99% |
| NutThread | 300-epoch PPO, trained 2026-08-01 | ~98% |

## ⚠️ Reproducing PegInsert

The PegInsert episodes were collected under **stock Forge dynamics**, but the Isaac Lab checkout
used for collection has `ForgeTaskPegInsert` modified for a separate study (higher and wider spawn,
plus a ±10° tilt of the peg *inside the gripper* to induce jams). Collection therefore overrode six
knobs back to their stock values:

```python
hand_init_pos            = [0.0, 0.0, 0.047]
hand_init_pos_noise      = [0.02, 0.02, 0.01]
hand_init_orn_noise      = [0.0, 0.0, 0.785]
held_asset_pos_noise     = [0.003, 0.0, 0.003]
fixed_asset_init_tilt_deg = 0.0
held_asset_tilt_deg       = 0.0
```

On an unmodified Isaac Lab these are already the defaults, so **run it as-is**. GearMesh and
NutThread were never modified. This matters: the 2026-06-30 PegInsert expert scores ~96% on stock
dynamics and **0/12** on the modified variant, and a policy trained on the modified variant plateaus
at a reward of 72.7 against 345.5 on stock — the same task name, two different problems.

## Loading

```python
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("angledusgar/forge-v1")
sample = ds[0]
sample["image"]      # (3, 256, 256)
sample["ft"]         # (6,)  force + torque
```

If video decoding fails with a `torchcodec` import error, the machine is missing FFmpeg shared
libraries; pass `video_backend="pyav"` or install ffmpeg.

## Known gaps

- **Arm joint angles are not included.** Only the two gripper joints were recorded; `state`
  follows the LIBERO convention, which is end-effector-based. Adding `joint_pos` requires
  re-collection and is planned for a future version.
- One shard was lost to a corrupted HDF5 file (a writer killed mid-write), costing at most 47
  PegInsert episodes out of ~1,010 collected for that task.
- Success rates are high by construction, so the failure set is small; see
  `forge-failure-v1` for the 51 failures and read its card before treating them as a balanced
  negative set.