charlesH777/grasp-anything-9040
08
Grasp Anything 2D — Checkpoint 9040 (grasp_rect 专用)
基于 NVIDIA LocateAnything-3B 的语言引导二维抓取模型,Phase 2 SFT 阶段产物。
任务
grasp_rect — 输出 (cx, cy, θ, width) 矩形抓取参数,按 RealVLG 固定 40 px depth 展开矩形。
评测结果(RealVLG 官方 652 样本)
- 评测协议:
evaluate_realvlg_grasp.py,fast 模式 - gAcccorrectedstrict: IoU>0.25 且 angle<30° (mod 180, 修正了官方 radian/degree bug)
- 对比 RealVLG-R1 GRPO Grasp 3B: seen gAcc +28.5%, similar +22.1%
使用方法
from locate_anything_service.model import LocateAnythingRuntime
from locate_anything_service.config import Settings
settings = Settings(model_id="charlesH777/grasp-anything-9040")
runtime = LocateAnythingRuntime(settings)
result = runtime.predict("image.jpg", "抓取红色杯子", mode="grasp_rect")训练配置
- 阶段: Phase 2 SFT (grasp_rect 专用)
- LLM LoRA: rank 32, target Qwen2.5-3B-Instruct
- 训练参数:
grasp_rect_task_enabled: true,contact_loss_enabled: false
