data-pipeline - SKILL.md Agent Skill

name: data-pipeline description: "Use for data recording and replay questions with dora. Triggers on: recording, replay, lerobot, data collection, dataset, rosbag, record data, play back, training data, HDF5, parquet, 数据记录, 数据回放, 数据收集, 训练数据" globs: ["/dataflow.yml", "/*.py"] source: "https://github.com/dora-rs/dora-lerobot"

Domain: Data Pipeline

Recording and replaying robot data with dora-rs

Overview

Dora supports data pipelines for:

Recording sensor data for training
Replaying recorded sessions
LeRobot integration for imitation learning
Dataset management

LeRobot Integration

Note: dora-lerobot is from a separate repository, not PyPI. Install from source:
git clone https://github.com/dora-rs/dora-lerobot
cd dora-lerobot
pip install -e dora_lerobot

Recording Data

nodes:
  # Camera
  - id: camera
    build: pip install opencv-video-capture
    path: opencv-video-capture
    inputs:
      tick: dora/timer/millis/33
    outputs:
      - image

  # Robot arm
  - id: arm
    build: pip install dora-rustypot
    path: dora-rustypot
    inputs:
      command: teleop/command
    outputs:
      - state
      - feedback

  # LeRobot recorder (install from dora-lerobot repo first)
  - id: recorder
    path: dora-lerobot-recorder
    inputs:
      image: camera/image
      state: arm/state
      action: teleop/command
    env:
      DATASET_NAME: my_robot_dataset
      EPISODE_INDEX: "0"

Replaying Data

nodes:
  # LeRobot replay (install from dora-lerobot repo first)
  - id: replay
    path: dora-lerobot-replay
    inputs:
      tick: dora/timer/millis/33
    outputs:
      - image
      - state
      - action
    env:
      DATASET_NAME: my_robot_dataset
      EPISODE_INDEX: "0"

  # Visualization
  - id: plot
    build: pip install dora-rerun
    path: dora-rerun
    inputs:
      image: replay/image

Data Collection Workflow

# 1. Start recording session
dora run record_dataflow.yml

# 2. Perform teleoperation
# 3. Press Ctrl+C to stop and save

# 4. Repeat for multiple episodes
EPISODE_INDEX=1 dora run record_dataflow.yml
EPISODE_INDEX=2 dora run record_dataflow.yml

# 5. Train policy
python train_policy.py --dataset my_robot_dataset

Available Hub Nodes

Node	Install	Purpose
dora-lerobot-recorder	From dora-lerobot repo	Record data
dora-lerobot-replay	From dora-lerobot repo	Replay data
llama-factory-recorder	`pip install llama-factory-recorder`	Record for LLM/VLM training
lerobot-dashboard	`pip install lerobot-dashboard`	Pygame recording interface
dora-rdt-1b	`pip install dora-rdt-1b`	VLA policy inference

Training Pipeline

1. Collect Demonstrations

# Run teleoperation dataflow
dora run record_dataflow.yml

# Mark episodes
# Press 'n' for new episode, 'f' for failed

2. Convert to LeRobot Format

# The recorder automatically saves in LeRobot format
# Dataset saved to: ~/.lerobot/datasets/<dataset_name>

3. Train Policy

# Using LeRobot CLI
python lerobot/train.py \
    --dataset my_robot_dataset \
    --policy diffusion \
    --output_dir outputs/my_policy

4. Deploy Policy

nodes:
  - id: camera
    path: opencv-video-capture
    inputs:
      tick: dora/timer/millis/33
    outputs:
      - image

  - id: policy
    build: pip install dora-rdt-1b
    path: dora-rdt-1b
    inputs:
      image: camera/image
    outputs:
      - action

  - id: robot
    path: dora-piper
    inputs:
      joint_positions: policy/action

Related Skills

hub-recording - Detailed recording node documentation
hub-robot - Robot control nodes
domain-robot - Robot control patterns
domain-vision - Visual data processing
hub-nodes - All pre-built nodes