Pi0☁云端推理部署

Pi0 是一个通用机器人策略基础模型，专为解决机器人学习中的数据稀缺、泛化能力差和鲁棒性不足等核心挑战而设计。借鉴大语言模型的训练方法，Pi0 通过大规模预训练掌握广泛的机器人操作技能，并能通过微调快速适应具体任务需求。该模型支持高效的数据利用与快速部署，在有限数据下也能实现良好性能，为机器人智能化提供了一种可扩展的解决方案。

1. 环境配置

在之前LeRobot安装及使用教程和Pi0 模型训练的基础上，Pi0的推理也建议使用云端部署的方式运行。

a. 我租赁的服务器是H20，在服务端（openpi环境）中运行：

cd /root/autodl-tmp/openpi_episode1_student
uv run scripts/serve_policy.py policy:checkpoint --policy.config=enpei_robot_demo_move_toy_low_mem_finetune --policy.dir=checkpoints/enpei_robot_demo_move_toy_low_mem_finetune/my_experiment/9999

启动服务器后，他会暴露一个服务端口（默认是 6006）：

warning: The `tool.uv.dev-dependencies` field (used in `packages/openpi-client/pyproject.toml`) is deprecated and will be removed in a future release; use `dependency-groups.dev` instead
INFO:root:Loading model...
INFO:2025-11-09 14:40:27,523:jax._src.xla_bridge:925: Unable to initialize backend 'rocm': module 'jaxlib.xla_extension' has no attribute 'GpuAllocatorConfig'
INFO:jax._src.xla_bridge:Unable to initialize backend 'rocm': module 'jaxlib.xla_extension' has no attribute 'GpuAllocatorConfig'
INFO:2025-11-09 14:40:27,524:jax._src.xla_bridge:925: Unable to initialize backend 'tpu': INTERNAL: Failed to open libtpu.so: libtpu.so: cannot open shared object file: No such file or directory
INFO:jax._src.xla_bridge:Unable to initialize backend 'tpu': INTERNAL: Failed to open libtpu.so: libtpu.so: cannot open shared object file: No such file or directory
INFO:absl:orbax-checkpoint version: 0.11.13
INFO:absl:Created BasePyTreeCheckpointHandler: use_ocdbt=True, use_zarr3=False, pytree_metadata_options=PyTreeMetadataOptions(support_rich_types=False), array_metadata_store=<orbax.checkpoint._src.metadata.array_metadata_store.Store object at 0x7f78fd908250>
INFO:absl:Restoring checkpoint from /root/autodl-tmp/openpi_episode1_student/checkpoints/enpei_robot_demo_move_toy_low_mem_finetune/my_experiment/9999/params.
INFO:absl:[thread=MainThread] Failed to get flag value for EXPERIMENTAL_ORBAX_USE_DISTRIBUTED_PROCESS_ID.
INFO:absl:[process=0] /jax/checkpoint/read/bytes_per_sec: 1.1 GiB/s (total bytes: 6.1 GiB) (time elapsed: 5 seconds) (per-host)
INFO:absl:Finished restoring checkpoint in 5.72 seconds from /root/autodl-tmp/openpi_episode1_student/checkpoints/enpei_robot_demo_move_toy_low_mem_finetune/my_experiment/9999/params.
INFO:root:Loaded norm stats from /root/autodl-tmp/openpi_episode1_student/assets/enpei_robot_demo_move_toy_low_mem_finetune/hou/demo_move_toy_openpi
INFO:root:Loaded norm stats from /root/autodl-tmp/openpi_episode1_student/checkpoints/enpei_robot_demo_move_toy_low_mem_finetune/my_experiment/9999/assets/hou/demo_move_toy_openpi
INFO:root:Creating server (host: autodl-container-9af746813b-5ccfe20e, ip: 172.17.0.8)
INFO:websockets.server:server listening on 0.0.0.0:6006
INFO:websockets.server:connection open
INFO:openpi.serving.websocket_policy_server:Connection from ('127.0.0.1', 53688) opened
INFO:openpi.serving.websocket_policy_server:Connection from ('127.0.0.1', 53688) closed
INFO:websockets.server:connection open
INFO:openpi.serving.websocket_policy_server:Connection from ('127.0.0.1', 56818) opened
INFO:openpi.serving.websocket_policy_server:Connection from ('127.0.0.1', 56818) closed

在AutoDL中使用它的端口转发服务（自定义服务）：

b. 在客户端（lerobot环境）进入Lerobot仓库代码，中安装openpi-client：

cd ./packages/openpi-client
pip install -e .

c. 启动机械臂进行归零，安装腕部相机和夹爪，查看相机ID:

(base) hou@hou-Ubuntu:~/workspace/lerobot_single_student$ conda activate lerobot
(lerobot) hou@hou-Ubuntu:~/workspace/lerobot_single_student$ python -m lerobot.episode_default_position --ip="localhost" --port=12345
INFO 2025-11-09 14:25:56 _position.py:33 Connected to EnpeiRobot controller at localhost:12345
INFO 2025-11-09 14:26:02 _position.py:46 Moving to default position, estimated time: 1.72s
INFO 2025-11-09 14:26:04 _position.py:50 Successfully moved to default position
(lerobot) hou@hou-Ubuntu:~/workspace/lerobot_single_student$ python -m lerobot.find_cameras opencv 

--- Detected Cameras ---
Camera #0:
  Name: OpenCV Camera @ /dev/video0
  Type: OpenCV
  Id: /dev/video0
  Backend api: V4L2
  Default stream profile:
    Format: 0.0
    Width: 640
    Height: 480
    Fps: 30.0
--------------------
Camera #1:
  Name: OpenCV Camera @ /dev/video2
  Type: OpenCV
  Id: /dev/video2
  Backend api: V4L2
  Default stream profile:
    Format: 0.0
    Width: 640
    Height: 480
    Fps: 30.0
--------------------
Camera #2:
  Name: OpenCV Camera @ /dev/video4
  Type: OpenCV
  Id: /dev/video4
  Backend api: V4L2
  Default stream profile:
    Format: 0.0
    Width: 640
    Height: 480
    Fps: 30.0
--------------------

Finalizing image saving...
Image capture finished. Images saved to outputs/captured_images

2. Pi0 推理

启动客户端，在终端中运行run.sh：

• host 服务器地址（这里因为用了 AutoDL 本地转发，所以是 localhost）
• port 服务器端口
• instruction 文本指令，保持和采集数据一致
• enpei_use_radian 需要使用弧度制

#!/bin/bash
python -m lerobot.test_openpi \
    --robot.ip_address="localhost" \
    --robot.port=12345 \
    --robot.type=enpei_follower \
    --robot.id=enpei_follower \
    --robot.cameras="{ handeye: {type: opencv, index_or_path: 4, width: 320, height: 240, fps: 30}, fixed: {type: opencv, index_or_path: 0, width: 320, height: 240, fps: 30}}" \
    --host=localhost \
    --port=6006 \
    --instruction="Put the toy to the white box" \
    --fps=30 \
    --enpei_use_radian=true

(lerobot) hou@hou-Ubuntu:~/workspace/lerobot_single_student$ ./run.sh 
INFO 2025-11-09 14:48:58 nt_policy.py:30 Waiting for server at ws://localhost:6006...
Connected to remote policy server at localhost:6006
Server metadata: {}
Using instruction: Put the toy to the white box
设置enpei follower机器人角度单位: 弧度
INFO 2025-11-09 14:48:58 follower.py:134 Connected to EnpeiRobot controller
INFO 2025-11-09 14:49:03 a_opencv.py:176 OpenCVCamera(4) connected.
INFO 2025-11-09 14:49:05 a_opencv.py:176 OpenCVCamera(0) connected.
INFO 2025-11-09 14:49:05 follower.py:156 enpei_follower EnpeiFollower connected.
Robot connected successfully
运行策略，目标FPS: 30
推理线程启动
等待第一个动作生成...
wait for frame: 0.0018558219999249559
wait for frame: 0.0018982179999511573
Loop duration: 1.4765s, Real FPS: 0.68, Action FPS: 33.87, Action count: 50, Action shape: (50, 7)
Loop duration: 1.4735s, Real FPS: 0.68, Action FPS: 33.93, Action count: 50, Action shape: (50, 7)
Loop duration: 1.4755s, Real FPS: 0.68, Action FPS: 33.89, Action count: 50, Action shape: (50, 7)
Loop duration: 1.4973s, Real FPS: 0.67, Action FPS: 33.96, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5040s, Real FPS: 0.66, Action FPS: 33.95, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5590s, Real FPS: 0.64, Action FPS: 33.95, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5400s, Real FPS: 0.65, Action FPS: 33.95, Action count: 50, Action shape: (50, 7)
Loop duration: 1.4911s, Real FPS: 0.67, Action FPS: 33.91, Action count: 50, Action shape: (50, 7)
Loop duration: 1.4938s, Real FPS: 0.67, Action FPS: 33.94, Action count: 50, Action shape: (50, 7)
Loop duration: 1.4853s, Real FPS: 0.67, Action FPS: 33.92, Action count: 50, Action shape: (50, 7)
Loop duration: 1.4957s, Real FPS: 0.67, Action FPS: 33.93, Action count: 50, Action shape: (50, 7)
Loop duration: 1.4930s, Real FPS: 0.67, Action FPS: 33.94, Action count: 50, Action shape: (50, 7)
Loop duration: 1.4914s, Real FPS: 0.67, Action FPS: 33.96, Action count: 50, Action shape: (50, 7)
Loop duration: 1.4878s, Real FPS: 0.67, Action FPS: 33.97, Action count: 50, Action shape: (50, 7)
Loop duration: 1.4911s, Real FPS: 0.67, Action FPS: 33.93, Action count: 50, Action shape: (50, 7)
Loop duration: 1.4914s, Real FPS: 0.67, Action FPS: 33.94, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5189s, Real FPS: 0.66, Action FPS: 33.96, Action count: 50, Action shape: (50, 7)
Loop duration: 1.4915s, Real FPS: 0.67, Action FPS: 33.94, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5207s, Real FPS: 0.66, Action FPS: 33.95, Action count: 50, Action shape: (50, 7)
Loop duration: 1.4865s, Real FPS: 0.67, Action FPS: 33.97, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5499s, Real FPS: 0.65, Action FPS: 33.95, Action count: 50, Action shape: (50, 7)
Loop duration: 1.6458s, Real FPS: 0.61, Action FPS: 33.93, Action count: 50, Action shape: (50, 7)
Loop duration: 1.6337s, Real FPS: 0.61, Action FPS: 33.98, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5200s, Real FPS: 0.66, Action FPS: 33.96, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5218s, Real FPS: 0.66, Action FPS: 33.93, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5184s, Real FPS: 0.66, Action FPS: 33.96, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5273s, Real FPS: 0.65, Action FPS: 33.96, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5247s, Real FPS: 0.66, Action FPS: 33.93, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5285s, Real FPS: 0.65, Action FPS: 33.95, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5576s, Real FPS: 0.64, Action FPS: 33.96, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5193s, Real FPS: 0.66, Action FPS: 33.93, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5519s, Real FPS: 0.64, Action FPS: 33.97, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5240s, Real FPS: 0.66, Action FPS: 33.94, Action count: 50, Action shape: (50, 7)
Loop duration: 1.6051s, Real FPS: 0.62, Action FPS: 33.93, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5251s, Real FPS: 0.66, Action FPS: 33.92, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5533s, Real FPS: 0.64, Action FPS: 33.94, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5564s, Real FPS: 0.64, Action FPS: 33.93, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5579s, Real FPS: 0.64, Action FPS: 33.92, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5550s, Real FPS: 0.64, Action FPS: 33.94, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5526s, Real FPS: 0.64, Action FPS: 33.96, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5507s, Real FPS: 0.64, Action FPS: 33.95, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5511s, Real FPS: 0.64, Action FPS: 33.96, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5557s, Real FPS: 0.64, Action FPS: 33.92, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5514s, Real FPS: 0.64, Action FPS: 33.95, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5539s, Real FPS: 0.64, Action FPS: 33.93, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5573s, Real FPS: 0.64, Action FPS: 33.96, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5767s, Real FPS: 0.63, Action FPS: 33.97, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5736s, Real FPS: 0.64, Action FPS: 33.93, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5553s, Real FPS: 0.64, Action FPS: 33.95, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5537s, Real FPS: 0.64, Action FPS: 33.97, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5514s, Real FPS: 0.64, Action FPS: 33.98, Action count: 50, Action shape: (50, 7)
Loop duration: 1.6261s, Real FPS: 0.61, Action FPS: 33.93, Action count: 50, Action shape: (50, 7)
Loop duration: 1.6378s, Real FPS: 0.61, Action FPS: 33.94, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5520s, Real FPS: 0.64, Action FPS: 33.93, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5588s, Real FPS: 0.64, Action FPS: 33.98, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5567s, Real FPS: 0.64, Action FPS: 33.97, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5534s, Real FPS: 0.64, Action FPS: 33.96, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5481s, Real FPS: 0.65, Action FPS: 33.97, Action count: 50, Action shape: (50, 7)
Loop duration: 1.6520s, Real FPS: 0.61, Action FPS: 33.95, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5514s, Real FPS: 0.64, Action FPS: 33.96, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5552s, Real FPS: 0.64, Action FPS: 33.95, Action count: 50, Action shape: (50, 7)
Loop duration: 1.5494s, Real FPS: 0.65, Action FPS: 33.93, Action count: 50, Action shape: (50, 7)
^C
Client stopped by user
推理线程结束

测试视频：

3. 小结

本文介绍了Pi0云端推理部署的完整流程。首先配置好服务器端环境并启动策略服务，然后在客户端安装所需依赖并连接机械臂与摄像头。实测下来，即使对于没见过的云宝机械臂的抓取放置成功率也能达到100%，体现了模型强大的泛化能力！