0
代码
代码
Issues
Pull Requests
讨论
Wiki
分析
0
BreezeWuBreezeWufeat(inference): support true distributed inference deployment
437e5f7a创建于 4月11日历史提交

robot_config

Unified robot configuration system for ros2_control and peripherals.

Overview

This package provides a unified configuration system for robot hardware that bridges:

  • ros2_control: For joint/motor control interfaces
  • Peripherals: For cameras and other devices (via existing ROS2 drivers)
  • tensormsg: For ML policy I/O contracts

The goal is to have a single source of truth for robot hardware configuration, eliminating duplication between different configuration systems.

Features

  • Single YAML configuration: Define ros2_control, cameras, and ML contracts in one file
  • Uses existing ROS2 camera drivers:
    • usb_cam for USB cameras (OpenCV-based)
    • realsense2_camera for RealSense D400 series
  • TF publishing: Automatic camera frame transform publishing
  • Calibration support: Standard ROS2 camera_info_manager integration
  • tensormsg integration: Contracts reference peripherals by name

Architecture

robot_config YAML (single source of truth)
        │
        ├───► ros2_control (joints/motors)
        │       └───► so101_hardware plugin
        │
        ├───► Camera drivers (existing ROS2 packages)
        │       ├───► usb_cam (USB cameras)
        │       └───► realsense2_camera (RealSense D400)
        │
        └───► tensormsg contracts (ML I/O)
                └───► PolicyBridge / EpisodeRecorder

Configuration Example

robot:
  name: so101_single_arm
  type: so101
  robot_type: so_101

  ros2_control:
    hardware_plugin: so101_hardware/SO101SystemHardware
    port: /dev/ttyACM0
    calib_file: $(env HOME)/.calibrate/so101_follower_calibrate.json
    reset_positions:
      "1": 0.0813
      "2": 3.7905

  peripherals:
    - type: camera
      name: top
      driver: opencv  # Uses usb_cam package
      index: 0
      width: 640
      height: 480
      fps: 30
      frame_id: camera_top_frame
      optical_frame_id: camera_top_optical_frame

    - type: camera
      name: wrist
      driver: realsense  # Uses realsense2_camera package
      serial_number: "12345678"
      width: 640
      height: 480
      fps: 30
      depth_width: 640
      depth_height: 480
      frame_id: camera_wrist_frame

  contract:
    observations:
      - key: observation.images.top
        topic: /camera/top
        peripheral: top  # References camera above
        image:
          resize: [480, 640]

Control Mode Configuration

The robot_config package supports dual control modes for different AI model requirements:

Available Control Modes

1. teleop Mode (Human Teleoperation)

Use for: Human teleoperation devices (leader arm, gamepad, VR device)

Characteristics:

  • Real-time direct control
  • Zero-latency passthrough (< 5ms)
  • Multiple input device support
  • Built-in safety filters (joint limits)

Configuration:

robot:
  default_control_mode: "teleop"

  control_modes:
    teleop:
      description: "Human teleoperation mode (direct control)"
      controllers:
        - joint_state_broadcaster
        - arm_position_controller
        - gripper_position_controller
      inference:
        enabled: false
        force_disable: true

  teleoperation:
    enabled: true
    active_device: "so101_leader"
    devices:
      - name: "so101_leader"
        type: "leader_arm"
        port: "/dev/ttyUSB0"
        calib_file: "$(env HOME)/.calibrate/so101_leader_calibrate.json"

Launch command:

# Teleop mode (with auto recording)
ros2 launch robot_config robot.launch.py robot_config:=so101_single_arm control_mode:=teleop record:=true

2. model_inference Mode (High-Frequency Position Control)

Use for: End-to-end imitation learning models (ACT, pi0, Diffusion Policy)

Characteristics:

  • High-frequency control (50-100Hz)
  • Low latency (1-3ms)
  • Direct topic-based position commands
  • Reactive, fluid movements

Configuration:

robot:
  default_control_mode: "model_inference"

  control_modes:
    model_inference:
      description: "High-frequency end-to-end control mode (ACT/pi0)"
      controllers:
        - joint_state_broadcaster
        - arm_position_controller
        - gripper_position_controller
      inference:
        enabled: true
        execution_mode: "distributed" # Or "monolithic" (single-machine zero-copy)
        model: so101_act

Launched controllers:

  • arm_position_controller (JointGroupPositionController)
  • gripper_position_controller (ForwardCommandController)

Command interface:

# Arm position commands
ros2 topic pub /arm_position_controller/commands std_msgs/msg/Float64MultiArray "data: [1.0, 2.0, 3.0, 4.0, 5.0]"

# Gripper position commands
ros2 topic pub /gripper_position_controller/commands std_msgs/msg/Float64MultiArray "data: [0.5]"

2. moveit_planning Mode (Trajectory Planning)

Use for: Planning-based models (VoxPoser, VLM, goal-conditioned policies)

Characteristics:

  • MoveIt integration (OMPL/Pilz planners)
  • Time-parameterized trajectories
  • Action-based execution with monitoring
  • Collision avoidance support

Configuration:

robot:
  default_control_mode: "moveit_planning"

  control_modes:
    moveit_planning:
      description: "MoveIt trajectory planning mode"
      controllers:
        - joint_state_broadcaster
        - arm_trajectory_controller
        - gripper_trajectory_controller

Launched controllers:

  • arm_trajectory_controller (JointTrajectoryController)
  • gripper_trajectory_controller (JointTrajectoryController)

Command interface:

# List available actions
ros2 action list

# Execute trajectory via action
ros2 action send_goal /arm_trajectory_controller/follow_joint_trajectory control_msgs/action/FollowJointTrajectory "{...}"

Overriding Control Mode at Runtime

Control mode can be overridden via command line:

# Use default mode from config file
ros2 launch robot_config robot.launch.py robot_config:=so101_single_arm

# Override to teleop mode
ros2 launch robot_config robot.launch.py robot_config:=so101_single_arm control_mode:=teleop

# Override to model_inference mode
ros2 launch robot_config robot.launch.py robot_config:=so101_single_arm control_mode:=model_inference

# Override to moveit_planning mode
ros2 launch robot_config robot.launch.py robot_config:=so101_single_arm control_mode:=moveit_planning

Mode Selection Decision Guide

What type of model are you using?
│
├─ Human teleoperation (leader arm, gamepad, VR)
│  └─ Use teleop mode
│     ├─ Real-time direct control
│     ├─ Zero-latency passthrough (< 5ms)
│     └─ Built-in safety filters
│
├─ End-to-end imitation learning (ACT, pi0, Diffusion)
│  └─ Use model_inference mode
│     ├─ Model outputs high-frequency position streams (50-100Hz)
│     ├─ Needs minimal latency (< 5ms)
│     └─ No trajectory planning required
│
└─ Planning-based (VoxPoser, VLM, goal-conditioned)
   └─ Use moveit_planning mode
      ├─ Model outputs sparse waypoints or goals
      ├─ Needs collision avoidance
      ├─ Requires MoveIt integration
      └─ Time-parameterized trajectories important

Complete Configuration Example

robot:
  name: so101_single_arm
  type: so101
  robot_type: so_101

  # Control mode management
  default_control_mode: "model_inference"  # Can be overridden via command line

  control_modes:
    teleop:
      description: "Human teleoperation mode (direct control)"
      controllers:
        - joint_state_broadcaster
        - arm_position_controller
        - gripper_position_controller
      inference:
        enabled: false
        force_disable: true

    model_inference:
      description: "High-frequency end-to-end control mode (ACT/pi0)"
      controllers:
        - joint_state_broadcaster
        - arm_position_controller
        - gripper_position_controller
      inference:
        enabled: true
        execution_mode: "distributed" # Or "monolithic" (single-machine zero-copy)
        model: so101_act

    moveit_planning:
      description: "MoveIt trajectory planning mode (VoxPoser/VLM)"
      controllers:
        - joint_state_broadcaster
        - arm_trajectory_controller
        - gripper_trajectory_controller

  # Unified joint configuration (DRY principle)
  joints:
    arm: ["1", "2", "3", "4", "5"]
    gripper: ["6"]
    all: ["1", "2", "3", "4", "5", "6"]

  # Hardware configuration
  ros2_control:
    hardware_plugin: so101_hardware/SO101SystemHardware
    port: /dev/ttyACM0
    calib_file: $(env HOME)/.calibrate/so101_follower_calibrate.json
    reset_positions:
      "1": 0.0813
      "2": 3.7905

  # Teleoperation configuration
  teleoperation:
    enabled: true
    active_device: "so101_leader"
    devices:
      - name: "so101_leader"
        type: "leader_arm"
        port: "/dev/ttyUSB0"
        calib_file: "$(env HOME)/.calibrate/so101_leader_calibrate.json"

  # Peripherals (cameras, sensors)
  peripherals:
    - type: camera
      name: top
      driver: opencv
      index: 0
      width: 640
      height: 480
      fps: 30

  # ML contract
  contract:
    observations:
      - key: observation.images.top
        topic: /camera/top
        peripheral: top
        image:
          resize: [480, 640]
    actions:
      - key: action
        topic: /arm_position_controller/commands  # Changes based on mode
        ros_type: std_msgs/msg/Float64MultiArray
        names: ["1", "2", "3", "4", "5", "6"]

How Mode Switching Works

  1. Configuration Phase:

    • robot.launch.py reads default_control_mode from YAML
    • Can be overridden via control_mode:=xxx command line argument
    • Validates mode exists in control_modes section

  2. Controller Spawning:

    • Only controllers listed in the selected mode are spawned
    • Ensures no controller conflicts (same joint can't be controlled by multiple controllers)

  3. Action Dispatch Integration:

    • action_dispatch node reads current mode from robot_config
    • Instantiates appropriate executor (TopicExecutor or ActionExecutor)
    • Provides unified API for upstream inference services

Troubleshooting Control Modes

Mode not switching

Problem: Command line override not taking effect

Solution: Ensure control_mode parameter is correctly spelled:

# Correct
ros2 launch robot_config robot.launch.py control_mode:=moveit_planning

# Incorrect (typo)
ros2 launch robot_config robot.launch.py control_mode:=moveit_planing

Controller not starting

Problem: Controllers fail to activate

Solution: Check controller configuration in so101_hardware/config/so101_controllers.yaml:

# Verify controller exists
ros2 control list_controllers

# Check controller configuration
cat src/so101_hardware/config/so101_controllers.yaml | grep -A 10 "arm_trajectory_controller"

Action server not available

Problem: FollowJointTrajectory action not found in moveit_planning mode

Solution: Ensure trajectory controllers are active:

# List active controllers
ros2 control list_controllers | grep trajectory

# Should see:
# arm_trajectory_controller[joint_trajectory_controller/JointTrajectoryController] active
# gripper_trajectory_controller[joint_trajectory_controller/JointTrajectoryController] active

For more details, see:

Usage

Launching the Robot

# Launch with real hardware
ros2 launch robot_config robot.launch.py robot_config:=so101_single_arm

# Launch with simulation
ros2 launch robot_config robot.launch.py robot_config:=so101_single_arm use_sim:=true

# Distributed inference — single-machine debug (Edge + Cloud co-located)
ros2 launch robot_config robot.launch.py robot_config:=so101_single_arm control_mode:=model_inference execution_mode:=distributed use_sim:=true cloud_local:=true

# Distributed inference — cross-machine (Device only launches Edge; Cloud runs separately on GPU server)
# Device side:
ros2 launch robot_config robot.launch.py robot_config:=so101_single_arm control_mode:=model_inference execution_mode:=distributed use_sim:=true
# GPU server (set same ROS_DOMAIN_ID):
# ros2 launch inference_service cloud_inference.launch.py policy_path:=/path/to/model device:=cuda

Validating Configuration

# Validate a robot config file with Python directly
python3 src/ros2/ros2_ws/src/robot_config/robot_config/scripts/validate_config.py \
    src/ros2/ros2_ws/src/robot_config/config/robots/so101_single_arm.yaml

Camera Drivers

USB Cameras (via usb_cam)

- type: camera
  name: usb_cam
  driver: opencv  # Uses usb_cam package
  index: 0  # USB device index (/dev/video0)
  width: 640
  height: 480
  fps: 30
  pixel_format: bgr8  # bgr8, rgb8, mono8, yuyv, etc.

安装 (Install):

# Ubuntu
sudo apt install ros-humble-usb-cam
# openEuler
sudo dnf install ros-humble-usb-cam

RealSense D400 Series (via realsense2_camera)

- type: camera
  name: rs_cam
  driver: realsense  # Uses realsense2_camera package
  serial_number: "12345678"  # Device serial (optional)
  width: 640
  height: 480
  fps: 30
  depth_width: 640
  depth_height: 480
  depth_fps: 30
  enable_pointcloud: false
  align_depth: false

安装 (Install):

# Ubuntu
sudo apt install ros-humble-librealsense2*
# openEuler
sudo dnf install ros-humble-librealsense2*

Camera Calibration

Camera intrinsics can be stored in the standard ROS2 location:

- type: camera
  name: top
  driver: opencv
  index: 0
  width: 640
  height: 480
  camera_info_url: file://$(env HOME)/.ros/camera_info/top_camera.yaml

Calibration files can be created using the standard ROS2 camera calibration tools:

ros2 run camera_calibration cameracalibrator \
  --size 8x6 \
  --square 0.024 \
  image:=/camera/top/image_raw

tensormsg Integration

The robot_config integrates with tensormsg contracts by allowing observations to reference peripherals by name:

# In robot_config
peripherals:
  - type: camera
    name: top
    width: 640
    height: 480

# In contract section
contract:
  observations:
    - key: observation.images.top
      topic: /camera/top
      peripheral: top  # Auto-fills width, height, fps from peripheral

When the contract is loaded, it will automatically include the camera metadata from the peripheral definition.

The old robot_interface package used LeRobot's Robot class directly. This package replaces it with:

  1. No LeRobot dependency in ROS2 layer: Uses ros2_control directly
  2. ros2_control native: Standard ROS2 hardware interface
  3. Existing ROS2 camera drivers: Uses usb_cam and realsense2_camera packages
  4. Single YAML configuration: All hardware defined in one place

The two example configurations from robot_interface have been manually migrated to:

  • config/robots/so101_single_arm.yaml (from single_arm_banana.yaml)
  • config/robots/so101_dual_arm.yaml (from dual_arms_pencil.yaml)

Troubleshooting

Camera not opening

Check USB permissions:

ls -l /dev/video*
sudo chmod 666 /dev/video0

Or add user to video group:

sudo usermod -a -G video $USER

RealSense camera not found

Install librealsense2:

# Ubuntu
sudo apt install librealsense2-utils librealsense2-dev
sudo apt install ros-humble-librealsense2*

# openEuler
sudo dnf install librealsense2-utils librealsense2-devel
sudo dnf install ros-humble-librealsense2*

Check camera is connected:

realsense-viewer

Calibration file not found

Make sure the path is correct and starts with file://:

camera_info_url: file:///home/user/.ros/camera_info/top.yaml

References

License

Apache-2.0