Attempt to be sample efficient in learning intuitive physics by treating the environment as a playground. Our uncertainty based approach learns 5x faster than vanilla SGD.
Here are a few results from the trained model.
Our model is probablistic, here are examples of a few rollouts,
This repository has a recreation of Chang et. al.'s neural physics engine, as well as our probablistic neural physics engine. The main driver file is
npe_main.py
.
Generate and train as,
python npe_main.py --gen_data --dataset PATH_TO_DATASET
python npe_main.py --train --dataset PATH_TO_DATASET --model PATH_TO_MODEL
python npe_main.py --model_simulation --model PATH_TO_MODEL
Of course, one can also just show the actual chipmunk simulation,
python npe_main.py --show_world
Just as a word of warning, this repo is still in the process of being cleaned.