README.md

---
license: apache-2.0
pipeline_tag: robotics
library_name: lerobot
tags:
- diffusion-policy
- flow-matching
- rectified-flow
- so101
- real-robot
- manipulation
- pick-and-place
- lerobot
datasets:
- wsagi/leisaac-real-pick-orange
base_model:
- wsagi/FlowHeads-DiffusionPolicy-PickOrange
---

# FlowHeads-DiffusionPolicy-Real-PickOrange — SO-101 真机抓橙子(sim-warm-start,FlowDP 真机 best)

真机 SO-101 pick-orange 的 **FlowDP(conv-UNet × rectified-flow head)best checkpoint**:从仿真 best [wsagi/FlowHeads-DiffusionPolicy-PickOrange](https://huggingface.co/wsagi/FlowHeads-DiffusionPolicy-PickOrange)(仿真榜 45.0%)warm-start,用 [wsagi/leisaac-real-pick-orange](https://huggingface.co/datasets/wsagi/leisaac-real-pick-orange)(30 集真机遥操演示)**续训 7000 步(≈4.5 epoch)**。

FlowDP (conv-UNet backbone × rectified-flow head) for real-world SO-101 pick-and-place: warm-started from our sim-best FlowDP checkpoint, fine-tuned for 7000 steps on 30 real teleoperation episodes. **Real-arm score even beats its own sim score** (56.7% vs 45.0% oranges/ep).

<video src="https://huggingface.co/wsagi/FlowHeads-DiffusionPolicy-Real-PickOrange/resolve/main/FlowDP-Real-PickOrange.mp4" controls muted loop></video>

## 真机结果 / Real-arm results(2026-07-13)

每轮场景放 3 只橙子,记每轮**放入盘中的橙子数**。Each round has 3 oranges on the table; we count oranges placed on the plate per round.

| 策略 / Policy | 轮次成绩 / Per-round | 均值 / Mean |
|---|---|---|
| **本模型(FlowDP simwarm 7000)/ this model** | 2, 2, 3, 1, 1, 2, 0, 3, 2, 1(10 轮) | **1.7 只/轮(56.7%)** |
| ACT simwarm 5000([wsagi/ACT-Real-PickOrange](https://huggingface.co/wsagi/ACT-Real-PickOrange)) | 2, 1, 1, 3, 2, 1, 0, 1, 2, 2(10 轮) | 1.5 只/轮 |
| FlowDP 从零 9000 / from-scratch | 能抓取,放置不稳 / grasps, unstable placement |  |

- 放入 ≥1 只的轮 9/10,两轮 3 只满放。**真机 56.7% 甚至超过同模型的仿真榜成绩 45.0%**。
- **sim-warm-start 再次变现**(与 ACT 线结论一致),且 FlowDP simwarm 反超 ACT simwarm(1.7 vs 1.5 只/轮)成为**真机全场 best**。

## 场景 / Scene

顶部相机视角(真机布置)/ top-camera view of the real setup:

![Top-Cam-DP](Top-Cam-DP.jpeg)

## 配方 / Recipe

| 项 | 值 |
|---|---|
| 架构 / Arch | FlowDP = lerobot DiffusionPolicy conv-UNet 骨干 × rectified-flow head(DDPM→flow 替换;horizon 16 / n_action_steps 8 / n_obs_steps 2) |
| init | [wsagi/FlowHeads-DiffusionPolicy-PickOrange](https://huggingface.co/wsagi/FlowHeads-DiffusionPolicy-PickOrange)(仿真 strict 20-round 45.0% best,step-9800) |
| 数据 / Data | [wsagi/leisaac-real-pick-orange](https://huggingface.co/datasets/wsagi/leisaac-real-pick-orange) 30 集 / 25,091 帧 @30fps |
| 续训 / Fine-tune | 7000 步(≈4.5 ep),lr 1e-4,batch 16,原生 480×640 |
| 观测 / Obs | `front` + `wrist` 相机 640×480 + 6 维关节位置 |

## 用法 / Usage

代码在 [vitorcen/LeIsaac-Training](https://github.com/vitorcen/LeIsaac-Training) 的 `FlowHeads/` 子包(`flowdp` 是自定义 policy type,需先注册再 `from_pretrained`):

```python
import sys; sys.path.insert(0, "<repo>/dependencies/FlowHeads")
import flowdp.configuration_flowdp  # registers type="flowdp"
from flowdp.modeling_flowdp import FlowDPPolicy

policy = FlowDPPolicy.from_pretrained("wsagi/FlowHeads-DiffusionPolicy-Real-PickOrange")
```

## 相关项目 / Related projects

- <https://github.com/vitorcen/isaaclab-experience> — 伞仓:多 VLA 基线训练/评测全记录
- <https://github.com/vitorcen/LeIsaac-Training> — LeIsaac:SO-101 PickOrange 训练 / benchmark / 真机脚手架