so100-act-training

Train an ACT (Action Chunking with Transformers) policy for the SO-100 robot arm using LeRobot on Google Colab, then run inference locally.

Overview

┌─────────────────┐      ┌──────────────┐      ┌─────────────────┐
│  Collect data   │ ──>  │  Train on    │ ──>  │  Run inference  │
│  (SO-100 + cams)│      │  Google Colab│      │  (local machine)│
└─────────────────┘      └──────────────┘      └─────────────────┘

Component	Details
Robot	SO-100 (6-DoF)
Policy	ACT – Action Chunking with Transformers
Framework	LeRobot
Training	Google Colab (GPU)
Inference	Any machine with USB connection to robot

Files

File	Purpose
train.py	Training script for Google Colab
inference.py	Inference script for local robot control

Prerequisites

• A HuggingFace account
• A LeRobot-format dataset uploaded to HuggingFace Hub
• Google account (for Colab)
• SO-100 robot arm (for inference)

1. Training (Google Colab)

Setup

1. Open Google Colab
2. Upload train.py or copy cells manually (each # %% = one cell)
3. Select a GPU runtime: Runtime > Change runtime type > GPU

Configure

Edit these variables in train.py (Cell 3):

DATASET_REPO_ID  = "your-hf-user/your-dataset"     # your HF dataset
CLEAN_EPISODES   = list(range(0, 60))              # episodes to train on
CORRUPT_EPISODES = [60, 61, 62]                    # episodes to skip

Edit these in Cell 7 (upload):

HF_USER    = "your-hf-user"
MODEL_NAME = "act_policy"

Run

Execute all cells in order. Training takes ~2–4 hours on a T4 GPU (100k steps).

The script automatically:

• Detects GPU and sets an appropriate batch size
• Installs LeRobot
• Loads and validates your dataset
• Trains the ACT policy
• Uploads the model to HuggingFace Hub

Hyperparameters

Key defaults (tuned for ~60 episodes of pick & place):

Parameter	Value	Notes
chunk_size	100	~3.3s lookahead at 30 Hz
training_steps	100,000
lr	1e-5	Increase to 5e-5 if loss stagnates
dropout	0.1	Increase to 0.2 if overfitting
temporal_ensemble_coeff	0.01	Lower = smoother actions

If the CLI training (Cell 6) fails, uncomment the fallback training loop in Cell 6b.

2. Inference (Local Machine)

Install

pip install lerobot

Run

python inference.py --model your-hf-user/act_policy

Options:

Flag	Default	Description
--model	(required)	HuggingFace model repo
--fps	30	Control loop frequency
--device	cpu	cpu or cuda

What it does

1. Downloads the trained policy from HuggingFace
2. Connects to the SO-100 robot
3. Runs a real-time control loop: observe → predict → act

Stop with Ctrl+C.

Troubleshooting

Problem	Solution
Loss stagnates	Increase lr to 5e-5
Overfitting	Increase dropout to 0.2, reduce training_steps
Colab timeout	Resume from last checkpoint in ./outputs/act_policy/
Jittery robot	Lower temporal_ensemble_coeff (e.g. 0.005)
FPS drops (inference)	Use --device cuda if GPU available, or reduce camera resolution

License

MIT