SO101 Arm ACT Training Guide Using SOLO CLI

This repository is made for anyone in the world who owns an SO101 robotic arm.

The datasets in this repository were recorded manually by me using the SO101 robot arm.

The goal is simple:

Download the dataset
Train the robot using SOLO CLI
Run inference
Make your SO101 robot perform the task automatically

Current tasks included:

Pick and Place
Liquid Pouring

This guide is written in a very simple copy-paste style so even beginners can follow it easily.

This guide shows how to:

Install SOLO CLI
Download your robotics dataset from Hugging Face
Train a robot policy model
Test the model
Understand which model to use

Everything is written in simple copy-paste steps.

What This Repository Helps You Do

If you have an SO101 robotic arm, you can use these datasets to teach your robot new skills using imitation learning.

The robot learns by watching demonstrations recorded in the dataset.

You do NOT need to manually code robot movement.

The AI model learns from the demonstrations and predicts robot actions automatically.

Available Datasets

Dataset	Description
`https://huggingface.co/datasets/omkarputti/SO101_pick_and_place`	Pick and Place task
`https://huggingface.co/datasets/omkarputti/SO101_liquid_pouring`	Liquid Pouring task

🔗 Hugging Face profile: https://huggingface.co/omkarputti

This guide uses only the ACT policy because it is reliable and works well for SO101 robot arm imitation-learning tasks like pick-and-place and liquid pouring.

System Requirements

Recommended:

Ubuntu Linux
NVIDIA GPU
CUDA installed
Python 3.10+

Example GPUs:

RTX 3050
RTX 3060
RTX 4060
RTX 4090

Step 1 — Install Basic Tools

Open terminal and copy-paste:

sudo apt update
sudo apt install git git-lfs python3 python3-pip python3-venv -y

Enable Git LFS:

git lfs install

Step 2 — Login to Hugging Face

Install Hugging Face CLI:

pip install -U huggingface_hub

Login:

huggingface-cli login

Paste your Hugging Face token.

Create token here: https://huggingface.co/settings/tokens

Choose: Write Access

Step 3 — Clone SOLO CLI

Go to Desktop:

cd ~/Desktop

Clone SOLO CLI:

git clone https://github.com/GetSoloTech/solo-cli

Open folder:

cd solo-cli

Step 4 — Create Python Environment

Create environment:

python3 -m venv solo-env

Activate environment:

source solo-env/bin/activate

Your terminal should now show:

(solo-env)

Step 5 — Install SOLO CLI

Install SOLO:

pip install -e .

Install training libraries:

pip install torch torchvision torchaudio

Install LeRobot:

pip install lerobot

Step 6 — Download Dataset

Move to a workspace folder:

mkdir -p ~/robot_datasets
cd ~/robot_datasets

Download Pick and Place dataset:

git clone https://huggingface.co/datasets/omkarputti/SO101_pick_and_place

OR download Liquid Pouring dataset:

git clone https://huggingface.co/datasets/omkarputti/SO101_liquid_pouring

Step 7 — Check Dataset Structure

Run:

cd S0101_pick_and_place
ls

You should see folders like:

meta/
data/
videos/

That means the dataset is valid.

Step 8 — Test Dataset Loading

Go back to SOLO CLI folder:

cd ~/Desktop/solo-cli

Open Python:

python

Paste:

from datasets import load_dataset

dataset = load_dataset("omkarputti/S0101_pick_and_place")
print(dataset)

If it prints dataset info — Everything works.

Exit Python:

exit()

🤖 What Happens During Training?

During training:

The AI watches robot demonstrations
Learns robot movement patterns
Learns how to move the arm
Learns how to complete the task

After training:

You get a trained robot AI model
The model can control the SO101 robotic arm

Step 9 — Start Training ACT Model

Train on Pick and Place Dataset

python -m lerobot.scripts.train \
  --dataset.repo_id=omkarputti/S0101_pick_and_place \
  --policy.type=act \
  --output_dir=outputs/act_pick_place \
  --job_name=act_pick_place_training

Train on Liquid Pouring Dataset

python -m lerobot.scripts.train \
  --dataset.repo_id=omkarputti/S0101_liquid_pouring \
  --policy.type=act \
  --output_dir=outputs/act_liquid_pouring \
  --job_name=act_liquid_pouring_training

Step 10 — Watch Training

During training you will see:

loss: 0.123
step: 100

Lower loss = better learning.

Training may take 30 minutes to several hours depending on your GPU.

Step 11 — Find Trained Model

After training:

cd outputs
ls

You will see:

act_pick_place/

Inside it:

checkpoints/

Your trained robot model is there.

What Is Inference?

Inference means using the trained AI model to control the real robot.

After training finishes:

The robot receives camera observations
The AI predicts robot arm actions
The robot performs the task automatically

Step 12 — Run Inference

python -m lerobot.scripts.eval \
  --dataset.repo_id=omkarputti/S0101_pick_and_place \
  --policy.path=outputs/act_pick_place/checkpoints/last/pretrained_model

Step 13 — Upload Trained Model to Hugging Face

Create model repo: https://huggingface.co/new-model

Then upload:

cd outputs/act_pick_place

Initialize git:

git init

Add remote:

git remote add origin https://huggingface.co/YOUR_USERNAME/YOUR_MODEL_NAME

Push:

git add .
git commit -m "upload model"
git push -u origin main

Important Tips

Tip 1 — Small Datasets

Your dataset is relatively small, which means:

Training is faster
Easier debugging
Good for learning

⚠️ But the model may overfit with limited data.

Tip 2 — More Data = Better Robot

To improve performance:

Record more trajectories
Add different lighting
Vary object positions
Add failures and recoveries

Robots learn from diversity.

🛠️ Common Problems

❌ Problem: CUDA not found

Check GPU:

nvidia-smi

Check PyTorch CUDA:

python -c "import torch; print(torch.cuda.is_available())"

Should print: True

❌ Problem: Hugging Face login failed

huggingface-cli login

Use your access token.

❌ Problem: Git LFS missing

sudo apt install git-lfs -y
git lfs install

❌ Problem: Dataset viewer missing

Sometimes Hugging Face needs 5–15+ minutes to index parquet files. This usually fixes itself automatically.

📚 Useful Commands

Command	Description
`nvidia-smi`	Check GPU
`source solo-env/bin/activate`	Activate environment
`deactivate`	Deactivate environment
`pip list`	Check installed packages

Official SOLO CLI Documentation

For setup errors or advanced commands: https://github.com/GetSoloTech/solo-cli

The repository contains:

Installation help
Training commands
Inference commands
Simulation setup
Troubleshooting
Dataset tools

🌍 Community Goal

This repository is designed to help SO101 robotic arm users:

Share datasets
Train robot policies
Improve imitation learning
Build open-source robotics projects
Make robotics easier for beginners

You can also contribute:

New datasets
Improved trajectories
Better camera angles
New robot tasks
Better trained models

🚀 Dataset Links

Task	Link
Pick and Place	https://huggingface.co/datasets/omkarputti/SO101_pick_and_place
Liquid Pouring	https://huggingface.co/datasets/omkarputti/SO101_liquid_pouring

Robot Hardware

SO101 Robotic Arm

The datasets were collected manually and uploaded to Hugging Face for imitation-learning training.

Name		Name	Last commit message	Last commit date
Latest commit History 17 Commits
ACT.png		ACT.png
Readme.Md		Readme.Md
so101.webp		so101.webp
solo-banner.png		solo-banner.png
training.png		training.png

Folders and files

Latest commit

History

Repository files navigation