---
license: apache-2.0
tags:
  - robotics
  - imitation-learning
  - dexflow
  - residual-mlp
  - dexterous-manipulation
library_name: pytorch
---

# DexFlow MLP Stage-2 v2 (M2 / M3 / M4)

Residual MLP Core4 Stage-2 variants on **plastic** and **pill** (30k, seed `20260910`, gate G2).

Org: [Humantwin](https://huggingface.co/Humantwin)  
Repo: [`Humantwin/dexflow-v2-stage2-tasks`](https://huggingface.co/Humantwin/dexflow-v2-stage2-tasks)

**Not ACT / DiT.** `DexFlowMLPPolicy` (`policy_variant=dexflow_mlp`).  
Visual backbone: ResNet18 → `Z_int[8×256]` → Residual MLP → action chunk `[30, 41]`.

Stage-1 init: `dexflow_mlp_stage1_core4_current_state_gate_20260910/stage1_step02000.ckpt`  
Code pack: `dexflow_v2_mlp_20260915.tar.zst`

## Variants

| Folder | Method | Head | Loss |
|---|---|---|---|
| `balanced_loss/` | M2 | unified | `0.5 L_body + 0.5 L_hand` |
| `split_head/` | M3 | shared trunk + BodyHead/HandHead | same balanced |
| `temporal_delta/` | M4 | same as M3 | balanced + `λ_Δ=0.1` hand temporal-delta |

Each variant has `plastic/` and `pill/`:

```text
{variant}/{task}/stage2.ckpt
{variant}/{task}/norm_stats.json
{variant}/{task}/gate_final.json
```

## Open-loop (train-set, G2 PASS @ 30k)

| Path | nRMSE | body RMSE | hand MAE |
|---|---:|---:|---:|
| `balanced_loss/plastic` | 0.0518 | 0.0053 | 2.48 |
| `balanced_loss/pill` | 0.0592 | 0.0065 | 3.74 |
| `split_head/plastic` | 0.0465 | 0.0048 | 2.48 |
| `split_head/pill` | 0.0501 | 0.0057 | 3.41 |
| `temporal_delta/plastic` | 0.0473 | 0.0048 | 2.37 |
| `temporal_delta/pill` | 0.0516 | 0.0058 | 3.55 |

## Load

```python
from deploy.real_robot.load_p0_policy import load_dexflow_policy
from dexflow.data.normalize import ActionStateNormalizer

model = load_dexflow_policy("split_head/plastic/stage2.ckpt", device="cuda")
normalizer = ActionStateNormalizer.load("split_head/plastic/norm_stats.json")
# image [B,1,3,480,640] float 0-1; state [B,41] MEAN_STD-normalized
chunk_n = model.predict_action_chunk({"image": img, "state": state_n})
action = normalizer.unnormalize_action(chunk_n)  # [B,30,41]
```

`from_checkpoint` restores `mlp_cfg.split_head` and `lambda_delta` from the ckpt.