Hugging Face's logo

Humantwin
/
dexflow-stage2-tasks

Robotics
PyTorch
imitation-learning
ACT
dexterous-manipulation
dexflow
Model card Files
xet
 Community
dexflow-stage2-tasks
File size: 3,169 Bytes
88d45e3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8e29c5a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
88d45e3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8e29c5a
 
 
 
 
 
88d45e3
 
 
 
 
8824a72
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: apache-2.0
tags:
  - robotics
  - imitation-learning
  - ACT
  - dexterous-manipulation
  - dexflow
library_name: pytorch
---

# DexFlow Stage-2 Checkpoints

Human→Robot DexFlow Stage-2 (Arena ACT + articulated HandFlow/ObjectFlow aux) checkpoints for Inspire-G1 tasks.

Org: [Humantwin](https://huggingface.co/Humantwin)

## Wave-2 P0 (2026-09-09) — recommended for bottle_upright real-robot

Same 10-ep **bottle_upright right crop cam30** robot set. P0-Ours additionally inits from Manus **plastic_bottle** Stage-1 (14 ep); S2 does not mix human RGB.

| ID | File | Init | S2 loss | Open-loop body / hand / J2 |
|----|------|------|---------|----------------------------|
| **P0-B0** | `p0s2_b0_bottle_cam30_crop/stage2.ckpt` | ImageNet | ACT only | **0.0085 / 4.69 / 1.49** |
| **P0-Ours** | `p0s2_ours_p0b_hf_bottle_cam30_crop/stage2.ckpt` | plastic S1 | ACT + 0.25 HF | **0.0083 / 4.83 / 1.72** |

Load:

```python
from dexflow.backbone.dexflow_act import DexFlowACTPolicy
model = DexFlowACTPolicy.from_checkpoint("p0s2_b0_bottle_cam30_crop/stage2.ckpt", stage="deploy")
model = model.strip_aux_for_deploy()
```

Deploy: publish 30/30, infer interval ≥1.05 s, strip aux, MEAN_STD `norm_stats.json` in the same folder.

---

## bottle_upright (right, crop cam30) — older 15k aux run

| Item | Value |
|---|---|
| File | `bottle_upright_right_crop/step_15000.ckpt` |
| Steps | 15000 / 30000 |
| Loss | L1 + 10·KL + 0.25·L_aux (HF+OF, contact off) |
| Init | Multi-task Stage-1 stage1_step06000 visual (+aux) |
| Data | bottle_upright right crop @ camera_30hz, 10 episodes |
| Norm | `bottle_upright_right_crop/norm_stats.json` (MEAN_STD) |

### Open-loop @ step 15000 (train-set, chunk step-0)

| Metric | Value |
|---|---:|
| body_rmse | 0.0134 rad |
| hand_mae | 9.62 (0–1000) |
| action_mse | 74.5 |

See `bottle_upright_right_crop/openloop_step15000/`.

### Load (DexFlow)

```python
import torch
ckpt = torch.load("bottle_upright_right_crop/step_15000.ckpt", map_location="cpu")
```

## Layout

```text
p0s2_b0_bottle_cam30_crop/          # Wave-2 P0-B0 @30k
  stage2.ckpt
  norm_stats.json
p0s2_ours_p0b_hf_bottle_cam30_crop/ # Wave-2 P0-Ours @30k
  stage2.ckpt
  norm_stats.json
bottle_upright_right_crop/
  step_15000.ckpt
  norm_stats.json
  openloop_step15000/
```

---

## Core4 current-state MLP G2 @ 25k (2026-09-11)

Four-task DexFlowMLP Stage-2 (action-only L1). Same 10 robot episodes / task. Same-set open-loop, not held-out SR.

| Task | File | nRMSE41 | body RMSE | hand MAE |
|---|---|---:|---:|---:|
| plastic / bottle_upright | `core4_mlp_current_state_g2_step25000/bottle_upright_right_crop/stage2_step25000.ckpt` | 0.050 | 0.0050 | 2.74 |
| pill | `core4_mlp_current_state_g2_step25000/pill_bottle_right_crop/stage2_step25000.ckpt` | 0.067 | 0.0059 | 4.16 |
| open_box_lid | `core4_mlp_current_state_g2_step25000/open_box_lid_both_crop/stage2_step25000.ckpt` | 0.054 | 0.0089 | 3.01 |
| realsense | `core4_mlp_current_state_g2_step25000/realsense_box_into_box_right_crop/stage2_step25000.ckpt` | 0.063 | 0.0062 | 3.63 |

Load: `DexFlowMLPPolicy.from_checkpoint(..., stage="deploy")` + same-folder `norm_stats.json`.