基于 Pi05 VLA 的玩偶分拣实战：从数据采集到实机推理

hydra:  # Hydra 配置文件保存目录，仅供参数检查使用
  run:
    dir: ./outputs/data_cvt_hydra_save/singlerun/${now:%Y%m%d_%H%M%S}  # 单次运行目录
  sweep:
    dir: ./outputs/data_cvt_hydra_save/multirun/${now:%Y%m%d_%H%M%S}  # sweep 时的根目录
    subdir: ${hydra:job.override_dirname}


# kuavo rosbag数据转换成lerobot-parquet格式的配置文件
# 请注意！！！
# 本分支的数转代码与 lerobot 0.5.2 同步
# 数据集格式使用的是 lerobot dataset v3.0
# lerobot地址：https://github.com/huggingface/lerobot

rosbag:
  mode: normal  # 可选: normal, resume, merge
  rosbag_dir: /home/lab/Desktop/lgn/data/rosbag/Sorting_dolls # rosbag文件存放的目录。normal / resume 模式需要
  target_dir: /home/lab/Desktop/lgn/data/lerobot_datasets/Sorting_dolls  # 输出父目录；最终数据集会生成在 target_dir/lerobot
  lerobot_dir_resume: /path/to/lerobot_resume  # resume 模式下，已有 lerobot 数据集的目录
  lerobot_dir_merge: 
    ["/path/to/lerobot1",
    "/path/to/lerobot2"]  # merge 模式下，要合并的多个 lerobot 数据集目录列表

dataset:
  platform_type: "5"  # 硬件平台类型，可选: "4pro", "5w" 或 "5"，默认使用 configs/platform/platform_config.yaml 中的 default 值，比赛默认4pro
  eef_type: leju_claw # 末端执行器类型，仿真选择：rq2f85, 真机可选：leju_claw,（夹爪） qiangnao，（灵巧手）
  which_arm: both  # 需要哪一只手臂的关节 + 图像数据，可选: left, right, both，注意图像数据会同时包含头部相机图像
 
  task_description: "Several red pony plush toys and yellow capybara plush toys are randomly placed in the center of the table. Sorting the toys, move the red pony plush toys to the left side of the table and the yellow capybara plush toys to the right side." # 任务描述，自定义

  train_hz: 10  # 训练数据的采样频率
  main_timeline: head_cam_h # 将哪个相机设为主相机，默认主相机来自：head_cam_h, 可选：wrist_cam_l, wrist_cam_r
  main_timeline_fps: 30 # 主相机的帧率，必须稳定，默认30帧
  sample_drop: 10 # 丢弃episode前后的10帧

  resize:
    width: 848  # 图像缩放宽度，注意请先查看（推荐浏览器foxglove工具）rosbag中图像的原始尺寸，避免放大或缩小带来模糊或拉伸形变等问题
    height: 480  # 图像缩放高度，注意请先查看（推荐浏览器foxglove工具）rosbag中图像的原始尺寸，避免放大或缩小带来模糊或拉伸形变等问题
    

    TrainConfig(
        # Change the name to reflect your model and dataset.
        name="pi05_sorting_dolls",
        # Here you define the model config -- In this example we use pi0 as the model
        # architecture and perform *full* finetuning. in the examples below we show how to modify
        # this to perform *low-memory* (LORA) finetuning and use pi0-FAST as an alternative architecture.
        model=pi0_config.Pi0Config(
            paligemma_variant="gemma_2b_lora",
            action_expert_variant="gemma_300m_lora",
            pi05=True,
        ),
        # Here you define the dataset you are training on. In this example we use the Libero
        # dataset. For your own dataset, you can change the repo_id to point to your dataset.
        # Also modify the DataConfig to use the new config you made for your dataset above.
        data=LeRobotKuavoDataConfig(
            repo_id="pi05_sorting_dolls",
            root="/home/kangkk/lgn/data/lerobot_data/Sorting_dolls/lerobot",
            base_config=DataConfig(
                # This flag determines whether we load the prompt (i.e. the task instruction) from the
                # ``task`` field in the LeRobot dataset. If set to True, the prompt will show up in
                # a field called ``prompt`` in the input dict. The recommended setting is True.
                prompt_from_task=True,
            ),
            extra_delta_transform=True,
        ),
        # Here you define which pre-trained checkpoint you want to load to initialize the model.
        # This should match the model config you chose above -- i.e. in this case we use the pi0 base model.
        weight_loader=weight_loaders.CheckpointWeightLoader("gs://openpi-assets/checkpoints/pi05_base/params"),
        # Below you can define other hyperparameters like the learning rate, number of training steps, etc.
        # Check the base TrainConfig class for a full list of available hyperparameters.
        num_train_steps=60_000,
        num_workers=8,
        wandb_enabled=True,
        batch_size=256,
        freeze_filter=pi0_config.Pi0Config(
            paligemma_variant="gemma_2b_lora", action_expert_variant="gemma_300m_lora"
        ).get_freeze_filter(),
    ),
    

# kuavo_server 统一启动配置 (configs/server/launcher.yaml)
#
# 用法（任选其一）：
#   python kuavo_server/launch.py openpi
#   python kuavo_server/launch.py groot
#   python kuavo_server/launch.py lingbotvla
#
# 启动器会读取下面对应 adapter 的 env 段（决定怎么进入环境：uv / conda / 当前 shell）
# 与 args 段（serve.py 的命令行参数），自动拼出原命令并 exec 运行。
#
# 规则：
#   - args 下的 key 直接对应 serve.py 的 --xxx 参数。
#   - 值为 true 的布尔会变成 --flag；值为 false / null / "" 的会被忽略，使用 adapter 内置默认。
#   - 路径可写绝对路径或 ${REPO_ROOT}/... 占位符（REPO_ROOT 自动指向本仓库根）。
#   - 每个 adapter 下方列出了**全部**可用参数；不想覆盖默认值的留空字符串 "" 即可。
#   - YAML 里的 key 是用户在命令行写的"短名"；若该短名 != 真实 adapter 名（如 groot != isaac_gr00t_n17），
#     用块内的 `adapter: <真实名>` 显式声明真实 adapter；不写时默认短名就是真实名（如 openpi）。

# 全局服务端参数（所有 adapter 共用，可被各 adapter 内的 args 覆盖）
common:
  host: "*"
  port: 5555
  api_token: ""

adapters:

  # ============================================================
  # OpenPI
  # ============================================================
  openpi:
    env:
      mode: uv                                                # uv | conda | none
      project: ${REPO_ROOT}/kuavo_model/external_models/openpi
    args:
      checkpoint: /home/lab/lgn/kuavo_learning_studio/checkpoints/pi05_sorting_dolls/pi05_sorting_dolls/20000                  # 必填：openpi checkpoint 目录
      policy_config_name: pi05_sorting_dolls                           # openpi 训练 config 名
      which_arm: both                                         # left | right | both（adapter 默认 right）
      model_repo_root: ""                                     # 留空 -> external_models/openpi
      execution_horizon: 0                                    # 0 -> 用 config.model.action_horizon
      device: ""                                              # "" | cuda | cuda:0 | cpu
      asset_id: ""                                            # "" -> auto-detect from checkpoint/assets/*

  # ============================================================
  # Isaac GR00T N1.7 
  # ============================================================
  groot:
    adapter: isaac_gr00t_n17
    env:
      mode: uv
      project: ${REPO_ROOT}/kuavo_model/external_models/gr00tn1d7
    args:
      checkpoint: /path/to/gr00t/checkpoint                   # 必填
      embodiment_tag: NEW_EMBODIMENT                          # 或 new_embodiment
      which_arm: both
      execution_horizon: 16                                   # receding horizon chunk size
      device: cuda
      strict: false                                           # true 时开启严格 IO 检查
      model_repo_root: ""                                     # 留空 -> external_models/gr00tn1d7

  # ============================================================
  # LingBot-VLA 
  # ============================================================
  lingbotvla:
    adapter: lingbot_vla
    env:
      mode: conda
      conda_env: lingbotvla
    args:
      checkpoint: /path/to/lingbot/hf_ckpt                    # 必填：hf_ckpt 目录
      training_config_path: /path/to/train_config.yaml        # 留空 -> configs/vla/real_load20000h.yaml
      qwen25_path: /path/to/Qwen2.5-VL                        # 必填：Qwen2.5-VL tokenizer/model 目录
      robot_norm_path: /path/to/robot_norm.json               # 必填：Kuavo norm stats JSON
      robot_config_path: ""                                   # 留空 -> kuavo_server/configs/lingbotvla/kuavo.yaml
      which_arm: both
      execution_horizon: 25                                   # -1 返回模型完整 chunk
      num_denoising_step: 10
      use_compile: false                                      # torch.compile
      use_fp32: false                                         # true 时用 fp32（默认 bf16）
      model_repo_root: ""                                     # 留空 -> external_models/lingbot-vla

env:  # 推理环境设置
  inference_env: real  # 选仿真还是真机 option: sim/real 
  platform_type: "5"  # 机器人型号，选项: "4pro", "5w" 或 "5"
  which_arm: both  # 控制左手、右手还是双手
  eef_type: leju_claw  # 夹爪或手型类型，仿真选择rq2f85，真机可选leju_claw（夹爪） qiangnao（灵巧手）
  ros_rate: 10  # 推理频率
  direct_to_wbc: true  # 是否开启direct to wbc模式。
  enable_action_interpolation: true # 开启direct to wbc后是否对动作进行插值，建议开启
  control_rate: 100  # WBC/插值后的实际控制指令频率

  

inference:  # 推理和评测设置
  go_bag_path: /home/leju/lgn/rosbag/A10-A15-L-H-01-TQ_09_01-P5_14-leju_claw-20260428170809-52-81c8a2-v003.bag # 推理时使用的rosbag路径
  policy_type: diffusion  # 要加载的模型类型,kuavo_server模式选择client, 选择client模式后下面的pretrained_path失去作用
  pretrained_path: /path/to/checkpoint # 要加载的权重路径。
  eval_episodes: 10  # 评测跑多少条
  max_episode_steps: 200  # 每条最多跑多少步

  task_prompt: "Several red pony plush toys and yellow capybara plush toys are randomly placed in the center of the table. Sorting the toys, move the red pony plush toys to the left side of the table and the yellow capybara plush toys to the right side."