Training Command Line - ACT (Recommended for Beginners)

Reference documentation

https://github.com/huggingface/lerobot/blob/46e19ae579f80ce66211afafd1c3c649c569131f/docs/source/act.mdx

https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/lerobot_train.py

https://github.com/huggingface/lerobot/blob/main/src/lerobot/configs/train.py

Why start with the ACT algorithm

ACT is the first model most recommended for training when playing with LeRobot. Its advantages are as follows:

• The model is very lightweight, with only eighty million learnable parameters

• Training converges very quickly, and inference is also very fast

• You can see results after training for one hour on a single GPU

• The ACT model download archive is about 200MB, which makes it very easy to store and transfer

• Collecting about 30 episodes of data for the dataset is basically enough

• It can be deployed for inference on an Ubuntu host, a Mac computer, a Windows computer, or even a Raspberry Pi

• The inference results on a real robot are still quite good, which is enough for simple tasks like grabbing, shaking hands, and placing a pen

• The ACT algorithm is already included in the basic environment of the LeRobot library, so no other libraries need to be installed

Command line

lerobot-train \
  --dataset.repo_id=Tommymy/lerobot_my_dataset_shake_hands \
  --dataset.root=~/lerobot_my_dataset_shake_hands \
  --dataset.revision=v0.1.0 \
  --dataset.streaming=false \
  --policy.type=act \
  --output_dir=~/output_lerobot_train/shake/act/ \
  --job_name=shake_act_a \
  --policy.device=cuda \
  --wandb.enable=true \
  --wandb.project=Lerobot_my_Project \
  --policy.push_to_hub=false \
  --steps=20000 \
  --batch_size=8

Command line reference

The line-continuation character \ can only have one space before it and no space after it

The parameters shown in red must be checked or modified before each run

Command line parameter	Description
--dataset.repo_id	The Repo_ID of the HuggingFace dataset
--dataset.root	The local path of the dataset
--dataset.revision	The dataset version, which was specified when uploading the dataset to HuggingFace
--dataset.streaming	The dataset is local, so it must be false, because the dataset is already local and does not need streaming reads
--dataset.split	Defaults to train, i.e. use all the data as the training set
--policy.type	The algorithm to train, such as act, smolvla, diffusion, pi0, wallx
--output_dir	The directory where the training output is saved
--job_name	The name of this training job
--policy.device	The compute device
--wandb.enable	Enable wandb visualization
--wandb.project	The wandb project name
--policy.push_to_hub	Upload the trained model to the HuggingFace cloud
--steps	The number of training steps
--batch_size	The amount of data fed in per step; if VRAM is insufficient, it should be reduced

Training process

The model archive is about 300MB