openpi-ex001-selectwaste

EX001 双臂移动机器人的「捡垃圾」策略权重($\pi_{0.5}$ 架构)。 Select-waste manipulation policy for the EX001 dual-arm mobile robot, in the $\pi_{0.5}$ architecture.

推理代码与完整部署教程:https://github.com/YangPaiMeng/openpi/tree/ex001-adapter 教程正文见 docs/ex001_selectwaste_zh.md

模型概况 / Overview

架构 $\pi_{0.5}$:PaliGemma 3B 视觉语言骨干 + 约 300M 动作专家
输入 3 路 RGB 图像(前视、左腕、右腕)+ 43 维机器人状态 + 语言指令
输出 48 步 × 28 维动作块(双臂末端位姿与夹爪、底盘速度、升降、头部转动)
精度 bfloat16
推理显存 约 24GB 起

策略输出的是以当前状态为锚点的相对动作,并按 horizon 逐步归一化。这两点与上游 openpi 的数据管线不同,因此必须配合上面链接里的适配代码使用,直接用上游 openpi 的标准推理脚本加载会得到错误的动作。

文件 / Files

文件 说明
params/ 模型权重(orbax ocdbt 格式,依赖精确的相对路径)
norm_stats.json 归一化统计量,缺少它无法正确推理
paligemma_tokenizer.model 分词器,避免运行时联网
SHA256SUMS 校验清单

下载后请务必校验(清单覆盖权重与统计量文件,不含本说明):

huggingface-cli download YangPaiMeng/openpi-ex001-selectwaste \
  --local-dir ./checkpoints/ex001-selectwaste
cd ./checkpoints/ex001-selectwaste && sha256sum -c SHA256SUMS

指令 / Prompts

本权重只在下面三条指令上训练过,使用其他措辞会明显降低表现:

Select the waste from the floor and put it into the waste bin by EVT robot.
Select the waste from the floor and put it into the waste bin by DVT robot and H gripper.
Select the waste from the floor and put it into the waste bin by DVT robot and G gripper.

安全提示 / Safety

这是一个直接驱动真实机械臂与移动底盘的策略。首次运行务必降低速度、缩小动作范围,并确保急停随时可用。本权重不适用于任何安全关键场景。

This policy directly commands real robot arms and a mobile base. Reduce speed and range of motion on first run, keep an emergency stop within reach, and do not use it in safety-critical settings.

License

This model is a Gemma Model Derivative: it was fine-tuned from the openpi $\pi_{0.5}$ base checkpoint, whose vision-language backbone is PaliGemma and therefore contains Gemma weights.

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms. A copy of the agreement is included in this repository as LICENSE_GEMMA.txt, and the required notice is in NOTICE.

Use of these weights is subject to the following restrictions, which are incorporated here as enforceable terms and which you must pass on to anyone you redistribute the weights or their functionality to. You must not use them:

Note that under the Gemma Terms of Use, "Distribution" includes making the model's functionality available as a hosted service via API or other remote means, not only shipping the weight files.

The inference code in the linked repository is licensed separately under Apache 2.0; this license section applies to the weights in this repository.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading