---
license: apache-2.0
language:
  - en
  - zh
tags:
  - robotics
  - grasping
  - vision-language-model
  - locate-anything
  - joint
  - multi-task
pipeline_tag: image-text-to-text
---

# Grasp Anything 2D — Checkpoint 2500 (joint 联合)

基于 NVIDIA [LocateAnything-3B](https://huggingface.co/nvidia/LocateAnything-3B) 的语言引导二维抓取模型,joint 联合训练产物。单 checkpoint 同时支持 `grasp_contact` 和 `grasp_rect` 两种任务。

## 任务

`grasp_contact` + `grasp_rect` 联合 — 单 checkpoint 双任务,共享 LLM backbone。

- contact: 输出平行夹爪两个接触点 `(x1, y1, x2, y2)`
- rect: 输出矩形抓取 `(cx, cy, θ, width)`

## 评测结果(RealVLG 官方 mini633)

| Split | n | gAcc_corrected_strict | mIoU_strict | center_err_median (px) |
|---|---|---|---|---|
| seen | 243 | **61.73%** | 55.07% | 11.11 |
| similar | 226 | **53.10%** | 46.85% | 15.00 |
| novel | 164 | 26.83% | 25.93% | 25.36 |

- 评测协议: `evaluate_realvlg_contact.py`,fast 模式
- **similar split gAcc 53.1% — 所有 checkpoint 中最高**(双任务正则带来更强泛化)
- 训练量仅 contact 专用的 1/8,性能却与 5x 训练量的 SFT checkpoint 持平

## 使用方法

```python
from locate_anything_service.model import LocateAnythingRuntime
from locate_anything_service.config import Settings

# joint checkpoint 两种模式都能跑
settings = Settings(model_id="charlesH777/grasp-anything-joint-2500")
runtime = LocateAnythingRuntime(settings)

# contact 任务
result = runtime.predict("image.jpg", "抓取红色杯子", mode="grasp_contact")

# rect 任务
result = runtime.predict("image.jpg", "抓取红色杯子", mode="grasp_rect")
```

## 训练配置

- 阶段: joint geometry (contact + grasp_rect 双任务联合)
- LLM LoRA: rank 32, target Qwen2.5-3B-Instruct
- 训练量: seen_contact_blocks = 19,941, seen_grasp_rect_blocks = 20,005
- contact_pair_weight: 1.0, grasp_rect_pose_weight: 1.0