Aloha

Experiments & demos with LeRobot Aloha robot arm

🖥️ System Information

• Operating System: Ubuntu 20.04 (via Windows WSL)
• GPU: NVIDIA GeForce RTX 4060, Driver Version 576.52, CUDA 12.9 【Train in RTX4090D 24G or AutoDL platform】
• Hardware: Logitech C920 Webcam, Xuanya hardware robotic arm based on Aloha （https://github.com/Xuanya-Robotics/lerobot）

📦 Features

• Simulation & real-world control
• Imitation learning & reinforcement learning
• Dataset collection & analysis

🎥 camera

• 介绍你使用的摄像头型号、分辨率、安装方式
• 示例：

  # 查看可用摄像头
  ls /dev/video*
  # 启动摄像头采集脚本
  python scripts/camera_setup.py
  
  

🧪 AutoDL Training Setup

🛠 Installation

git clone https://github.com/yourusername/aloha-project.git
cd aloha-project
# install dependencies etc.


````markdown
#

This section documents the training environment setup and execution on the **AutoDL platform**, including conda environment setup, code installation, Hugging Face integration, and training script usage.

---

### 🔧 Environment Setup

1. **Initialize Conda**
```bash
conda env list
conda activate base
conda init

Restart the terminal after initializing conda.

2. Create and activate environment

conda create -y -n lerobot python=3.10
conda activate lerobot

3. Enable Network Turbo (AutoDL Specific)

source /etc/network_turbo
# Output: Successfully enabled. Note: Academic use only, no stability guaranteed.

---

📦 Project Setup

4. Clone project repositories

git clone https://github.com/Xuanya-Robotics/lerobot.git
cd lerobot
git clone https://github.com/Xuanya-Robotics/Alicia_duo_sdk.git Alicia_duo_sdk

5. Install dependencies

conda install ffmpeg -c conda-forge
pip install -e .

---

🔐 Hugging Face Configuration

1. Set Git Credential Helper

git config --global credential.helper store

2. Login using token

hf auth login --token <your_token>
# Example:
# hf auth login --token hf_..................... 

After login, the token will be saved locally. You may add --add-to-git-credential if needed.

3. Set Hugging Face username environment variable

HF_USER=$(hf auth whoami | head -n 1) && echo $HF_USER
# Example output: Enstar07

---

🏋️ Training Examples

📡 Training from Hugging Face cloud dataset

python lerobot/scripts/train.py \
  --dataset.repo_id Enstar07/demo_dataset3 \
  --policy.type act \
  --output_dir outputs/train/act_alicia_duo_model_final \
  --job_name alicia_duo_act_training_final \
  --policy.device cuda \
  --batch_size 32 \
  --steps 2000 \
  --save_freq 500 \
  --eval_freq 500 \
  --log_freq 100

⚠️ Cloud-based dataset training may suffer from network delays. For large datasets, consider using FileZilla to upload local data directly to AutoDL.

---

💾 Training from Local Dataset (uploaded via FileZilla)

python lerobot/scripts/train.py \
  --dataset.repo_id /root/lerobot/data/zhuiPick3 \
  --policy.type act \
  --output_dir outputs/train/act_alicia_duo_model_final \
  --job_name alicia_duo_act_training_final \
  --policy.device cuda \
  --batch_size 32 \
  --steps 2000 \
  --save_freq 500 \
  --eval_freq 500 \
  --log_freq 100

---

🚀 Full-scale Training (successful)

python lerobot/scripts/train.py \
  --dataset.repo_id /root/lerobot/data/zhuiPick3 \
  --policy.type act \
  --output_dir outputs/train/act_alicia_duo_model_final \
  --job_name alicia_duo_act_training_final \
  --policy.device cuda \
  --batch_size 64 \
  --steps 20000 \
  --save_freq 1000 \
  --eval_freq 1000 \
  --log_freq 500

---

📤 Model Export & Deployment

• After training completes, use FileZilla to download the trained model directory to your local system.
• Update the deployment script path:

# In dp_inference.py
ckpt_path = "/home/zexuan/Robot/xuanArm/system2/lerobot/outputs/train/filezilla/020000/pretrained_model"

• Then run deployment:

cd examples/
python dp_inference.py

---

---

 ## 🧪 AutoDL Training Notes