🚀 XLeRobot VLA Development Tutorial

May 12, 2026 · View on GitHub

Build a full robot control pipeline from data → training → real-world deployment using:

  • Dexbotic toolbox
  • XLeRobot (v0.4.0, dual-wheel version)

👉 This guide is designed to help you run the entire pipeline end-to-end from scratch

Pipeline Overview

Before we begin, here’s what you will build:

flowchart LR
    A[Data Collection]
    B[Data Conversion]
    C[Dexbotic Policy Training]
    D[Real-World Deployment]

    A --> |"LeRobot"| B --> |"DexData format"| C --> |"Policy Model"| D

Inference Architecture

The XLeRobot inference system follows a decoupled, three-tier architecture:

sequenceDiagram
    participant Robot
    participant Bridge
    participant Policy

    Robot->>Bridge: Observations (gRPC)
    Bridge->>Policy: HTTP Request
    Policy-->>Bridge: Actions
    Bridge-->>Robot: Actions (gRPC)

🧩 Phase 0: Preparation

Follow the official XLeRobot documentation and complete:

  1. Dual-wheel assembly
  2. Install LeRobot
  3. Dataset collection

🧩 Phase 1: Data Conversion (LeRobot → DexData)

🧠 What happens during the conversion?

  • Align images, states, actions, and prompts
  • Synchronize timestamps
  • Split data into episodes

1. Input Data Structure (LeRobot)

xlerobot-data/
└── <task_name>/                       # e.g. pick_longan_to_box
    ├── data/
    │   └── chunk-000/
    │       ├── file-000.parquet       # State/action data
    │       ├── file-001.parquet
    │       └── ...
    ├── meta/
    │   ├── episodes/
    │   │   └── chunk-000/
    │   │       ├── file-000.parquet   # Episode metadata
    │   │       ├── file-001.parquet
    │   │       └── ...
    │   ├── info.json                  # Dataset info (fps, features, paths)
    │   ├── stats.json
    │   └── tasks.parquet              # Task metadata (prompts)
    └── videos/
        ├── observation.images.head/
        │   └── chunk-000/
        │       ├── file-000.mp4
        │       └── ...
        ├── observation.images.wrist_left/
        │   └── chunk-000/
        │       ├── file-000.mp4
        │       └── ...
        └── observation.images.wrist_right/
            └── chunk-000/
                ├── file-000.mp4
                └── ...

2. Output Data Structure (DexData)

xlerobot_dexdata/
├── jsonl/
│   └── <task_name>/
│       ├── episode_00000.jsonl
│       ├── episode_00001.jsonl
│       └── ...
└── video/
    └── <task_name>/
        ├── episode_00000_head.mp4
        ├── episode_00000_wrist_left.mp4
        ├── episode_00000_wrist_right.mp4
        ├── episode_00001_head.mp4
        └── ...

3. Run Conversion

python convert_xlerobot_to_dexdata.py \
  --lerobot_dir /path/to/xlerobot-data/<task_name> \
  --output_dir /path/to/xlerobot_dexdata

🧩 Phase 2: Policy Training

1. Dataset Registration

Register your dataset in dexbotic/data/data_source/xlerobot_<task>.py:

from dexbotic.data.data_source.register import register_dataset

XLEROBOT_DATASET = {
    "<task_name>": {
        "data_path_prefix": "/path/to/dexdata/video",
        "annotations": "/path/to/dexdata/jsonl/<task_name>",
        "frequency": 1,
    },
}

meta_data = {
    'non_delta_mask': [5, 11, 12, 13, 14, 15],  # Non-delta action indices
    'periodic_mask': None,
    'periodic_range': None
}

register_dataset(XLEROBOT_DATASET, meta_data=meta_data, prefix='xlerobot')

Important: The non_delta_mask [5, 11, 12, 13, 14, 15] corresponds to: Grippers (5, 11), Head Motors (12, 13), and Wheel Velocities (14, 15).

2. Create an Experiment Config

Create a config file:

playground/train_xlerobot_<your_task>.py

Required settings:

# Dataset
DM0DataConfig.dataset_name = "your_task"
DM0DataConfig.num_images = 3
DM0DataConfig.images_keys = ["images_1", "images_2", "images_3"]  # head, wrist_left, wrist_right

# Training
DM0TrainerConfig.output_dir = "..."  # checkpoint/output directory

# Model
DM0ModelConfig.model_name_or_path = "..."  # pretrained model path

# Inference
DM0InferenceConfig.model_name_or_path = "..."  # finetuned checkpoint
DM0InferenceConfig.port = 7891
DM0InferenceConfig.non_delta_mask = [5, 11, 12, 13, 14, 15]  # must match dataset
DM0InferenceConfig.action_dim = 16  # XLeRobot action space

3. Training

cd dexbotic
deepspeed playground/train_xlerobot_<your_task>.py --task train

🧩 Phase 3: Real-World Deployment

Terminal 1 — VLA Policy Server (GPU server / workstation):

cd dexbotic
python playground/train_xlerobot_<your_task>.py --task inference

Terminal 2 — Bridge (on robot controller / edge device):

cd dexbotic
python hardware/xlerobot/bridge.py \
  --port 8080 \
  --vla_url http://127.0.0.1:7891 \
  --prompt "TASK_PROMPT" \
  --show_images

Terminal 3 — Robot Client (on robot controller / edge device):

cd lerobot
export SVT_LOG=1
python -m lerobot.async_inference.robot_client \
    --robot.type=xlerobot_2wheels \
    --robot.id=xler \
    --robot.port1=/dev/tty.usbmodem<SERIAL_1> \
    --robot.port2=/dev/tty.usbmodem<SERIAL_2> \
    --robot.cameras='your_camera_config_json' \
    --server_address=127.0.0.1:8080 \
    --policy_type=remote \
    --pretrained_name_or_path="none" \
    --actions_per_chunk=50 \
    --chunk_size_threshold=0.5 \
    --aggregate_fn_name=weighted_average \
    --fps=30

For detailed parameter descriptions, refer to the LeRobot Asynchronous Inference Documentation.

Important: Remember to add 'xlerobot_2wheels' to the SUPPORTED_ROBOTS list in lerobot/async_inference/constants.py.

🔥 FAQ

Q1: Cameras not working

  • Verify devices are detected:

    v4l2-ctl --list-devices

  • Check if /dev/video* exists

  • Ensure the device is not occupied:

    lsof /dev/video0

  • Fix permission issues:

    sudo chmod 666 /dev/video0

Q2: Robot fails to move

  • The bridge is not connected to the VLA server

Q3: Bad Inference

  • Incorrect camera mapping
  • Prompt mismatch (MOST COMMON)
  • Incorrect non_delta_mask or action_dim
  • Incorrect model checkpoint
  • Camera moved (camera extrinsics changed)

Q4: High latency

  • Increase GPU performance
  • Reduce resolution
  • Reduce cameras

Q5: Robot jittery (unstable or shaky motion)

  • Control frequency mismatch
  • FPS mismatch (sensor vs. control loop)
  • Improper or missing action aggregation

Q6: Optimization

  • Collect more diverse data
  • Improve data quality via noise filtering and outlier detection
  • Ensure temporal alignment (action ↔ observation)
  • Consider using the DM0 model

Demos

Following the above instructions, we conducted a Collect Longans into the Box task on our XLeRobot robot, and provide a successful demo video below:

📹 View Demo Video: Collect Longans into the Box