(use with caution. this repo is still a work in progress)
A tiny World Action Model from scratch for robot learning.
mini-WAM extends the educational philosophy of mini-VLA.
Where mini-VLA learns:
image_t + language + state_t → action_t
mini-WAM learns:
image_t + language + state_t → action_t + state_t+1 + image_t+1
A VLA learns to act. A WAM learns to act while predicting what its action will do to the world.
pip install -r requirements.txt
python scripts/collect_data.py --env toy --num_episodes 100
python scripts/train.py \
--env toy \
--model_type wam \
--action_head mlp \
--predict_state \
--device cpu