BREE

BREE dual-arm system operating guideBREE 双臂系统操作指南

This document assumes that all commands are run from the LeRobot repository root:本文档默认所有命令都从 LeRobot 仓库根目录运行:

bash
cd lerobot
conda activate lerobot

If you are not using a Conda environment, first confirm that the current Python can run python -c "import lerobot".如果使用的不是 Conda 环境,也应先确认当前 Python 能够执行 python -c "import lerobot"。

Follow the workflow按流程开始

1. Complete workflow overview1. 完整流程概览

Choose the workflow that matches your current task. Recalibration is not required every time.根据当前工作内容选择对应流程,不需要每次都重新校准。

First installation or hardware replacement首次安装或更换硬件

  1. Identify the serial ports and grant access permissions.确认串口并赋予权限。
  2. Calibrate all four arms.校准四条机械臂。
  3. Confirm that the arms can safely return to the H pose.确认机械臂能够安全回到 H 位。
  4. Confirm the device indices of all three cameras.确认三个相机的设备编号。
  5. Adjust the global camera and both wrist cameras.调整全局相机和两个腕部相机。
  6. Regenerate camera reference files only when the installation baseline needs to be redefined.只有在需要重新定义安装基准时,才重新生成相机参考文件。

Daily workflow after each startup每次开机后的日常流程

  1. Grant access permissions to the serial ports.给串口赋予权限。
  2. Check the serial ports and camera indices.核对串口和相机编号。
  3. Check the arms at the H pose.检查机械臂 H 位。
  4. Use the existing reference images to check the positions of all three cameras.用已有参考图检查三个相机的位置。
  5. Proceed with teleoperation, data collection or model evaluation.进行遥操作、数据采集或模型评估。

Training and evaluation workflow训练与评估流程

  1. Prepare task scene reference images.准备任务场景参考图。
  2. Collect data with lerobot-record.使用 lerobot-record 采集数据。
  3. Train a model with lerobot-train.使用 lerobot-train 训练模型。
  4. Evaluate with lerobot-record, the RDT client or the OpenPI client.使用 lerobot-record、RDT 客户端或 OpenPI 客户端进行评估。
  5. Write all run results to outputs/ under the repository root, rather than to BREE/.所有运行结果统一写入仓库根目录下的 outputs/,不要写入 BREE/。

2. Directory guide2. 目录说明

Path路径Purpose用途
BREE/robot_calibration/Arm calibration, H-pose verification and wrist repeatability verification scripts机械臂校准、H 位验证和腕部重复性验证脚本
BREE/camera_calibration/Camera-view, white-balance and reference-image alignment scripts相机画面、白平衡和参考图对齐脚本
BREE/camera_reference/Camera installation reference files currently in use当前使用的相机安装参考文件
BREE/model_inference/RDT dual-arm evaluation entry pointRDT 双臂评估入口
BREE/openpi_eval/OpenPI code, environment, models, server and clientOpenPI 代码、环境、模型、服务端和客户端
outputs/Data processing, training, evaluation videos and check results数据处理、训练、评估视频和检查结果

The repository already contains the following camera installation references:当前仓库内已经存在以下相机安装参考:

text
BREE/camera_reference/
├── global_image/
│   ├── reference.jpg
│   └── reference.json
└── overlay/
    ├── left_wrist/
    │   ├── reference.jpg
    │   └── reference.json
    └── right_wrist/
        ├── reference.jpg
        └── reference.json

Daily checks should read these files. Do not overwrite them by running save-reference.日常检查应读取这些文件,不要运行 save-reference 覆盖它们。

12. Output locations12. 输出位置

Place all newly generated results under outputs/ at the repository root:所有新生成结果都放在仓库根目录的 outputs/ 下:

Result结果Recommended location推荐位置
Camera alignment checks相机对齐检查outputs/camera_check/
Arm calibration verification机械臂校准验证outputs/robot_calibration/
Model training模型训练outputs/train/
RDT language embeddingsRDT 语言向量outputs/rdt_lang/
RDT evaluation videosRDT 评估视频outputs/rdt_eval_videos/

BREE/ stores code, fixed reference files and OpenPI runtime dependencies. It does not store the results generated by individual runs.BREE/ 用于保存代码、固定参考文件和 OpenPI 运行依赖,不用于保存每次运行产生的结果。