Step-by-Step Guide: Train a Lerobot Robotic Arm from Scratch on GPUFree

Preparation

Ensure the dataset directory exists, e.g. ~/.cache/huggingface/lerobot/mytest/so100_test. The system disk is limited to 30 GB; using a larger data disk is recommended.

Rent a GPUFree cloud instance

Register or log in at https://www.gpufree.cn/market.

Select an RTX 4090 instance (≈ ¥1.38 per hour).

Choose an Ubuntu 20.04 or 22.04 image with Miniconda or PyTorch pre‑installed.

Recharge about ¥10 and create the instance.

Start the instance with GPU enabled for training; use the “no‑GPU” mode for non‑training operations.

Web UI for remote training

The official Lerobot image on GPUFree provides a Flask web service that handles dataset upload, training launch and model download.

One‑click upload of a local dataset folder.

Select training algorithm (ACT, Pi0.5, Smolval, etc.) and start training.

Automatic shutdown of the instance after training to avoid extra charges.

Download the zipped model directly from the browser.

Start the Flask service

Create a JupyterLab notebook or a Python file named remote_train.ipynb (or remote_train.py) and run the following code. The service listens on port 7001, which GPUFree maps to a public address.

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""LeRobot 远程训练平台"""
import os, time, subprocess, threading, shutil, re
from flask import Flask, request, jsonify, send_file, render_template_string
from werkzeug.utils import secure_filename

app = Flask(__name__)

UPLOAD_FOLDER = './upload_temp'
DOWNLOAD_FOLDER = './output'
DOWNLOAD_TEMP_FOLDER = './download_temp'
MAX_CONTENT_LENGTH = 10*1024*1024*1024  # 10 GiB

app.config['UPLOAD_FOLDER'] = UPLOAD_FOLDER
app.config['DOWNLOAD_FOLDER'] = DOWNLOAD_FOLDER
app.config['MAX_CONTENT_LENGTH'] = MAX_CONTENT_LENGTH

def ensure_directories():
    for folder in [UPLOAD_FOLDER, DOWNLOAD_FOLDER, DOWNLOAD_TEMP_FOLDER]:
        if not os.path.path.exists(folder):
            os.makedirs(folder, exist_ok=True)

# ... (Flask routes for /list_datasets, /start_training, /download_model) ...

Upload the collected dataset

In the UI click “Upload”, select the hidden folder ~/.cache/huggingface/lerobot/.../so100_test (press Ctrl+H to show hidden files). Upload speed is about 3 MiB/s; the data is stored on the instance under /root/data/upload_temp.

Start training

Select the uploaded dataset, choose an algorithm (e.g., ACT), optionally adjust parameters, and click “Start Training”. Real‑time logs are displayed. When training finishes the model is zipped and placed in /root/data/download_temp. Enable “auto‑shutdown” to terminate the instance automatically.

Download the model

The right‑hand side of the UI lists the zipped model file. Click to download, then unzip locally, for example:

lerobot-record \
  --robot.type=so100_follower \
  --robot.port=/dev/ttyACM1 \
  --robot.cameras="{ up: {type: opencv, index_or_path: /dev/video10, width: 640, height: 480, fps: 30}, side: {type: intelrealsense, serial_number_or_name: 233522074606, width: 640, height: 480, fps: 30}}" \
  --robot.id=my_awesome_follower_arm \
  --display_data=false \
  --dataset.repo_id=${HF_USER}/eval_so100 \
  --dataset.single_task="Put lego brick into the transparent box" \
  --policy.path={your_unzipped_directory}

Key implementation details

The Flask service defines the following routes (simplified): /list_datasets (GET) – returns a JSON list of uploaded datasets with validation. /start_training (POST) – receives JSON with dataset, algorithm, and optional parameters, launches the training subprocess, and stores process handles and logs. /download_model (GET) – packages the latest output directory into a zip archive and returns the file.

Utility functions ensure_directories, extract_output_dir, and pack_model handle directory creation, parsing of the training command for the --output_dir argument, and creation of a timestamped zip archive.

Summary

Using a GPUFree RTX 4090 instance reduces training time dramatically compared with local CPU training. The web UI and the ~500‑line Flask service allow dataset upload, training launch, automatic shutdown, and model download without manual Linux configuration. The entire workflow—from preparing the dataset to running lerobot-record with the downloaded policy—fits within a few hundred lines of Python and a few hundred yuan of cloud cost.