Propagation Networks for Model-based Control Under Partial Observation

Yunzhu Li, Jiajun Wu, Jun-Yan Zhu, Joshua B. Tenenbaum, Antonio Torralba, Russ Tedrake

ICRA 2019 [website] [paper] [video]

Demo

Simulation

Newton's Cradle

Rollout from the learned model, where we show the ground truth in transparent. The discrpancy is almost not noticeable, which indicates that our model can successfully handle instantaneous propagation of forces.

Control

Rope Manipulation

Shake a rope to match a target configuration shown in transparent, where we are only allowed to apply forces to the top 2 particles at the free end. Yellow arrow indicates the applied force.

Box Pushing

Push a pile of boxes to a target configuration shown in transparent. Note that we assume partial observability in this example where we are viewing the scene from the top and only red boxes are visible.

Installation

This codebase is tested with Ubuntu 16.04 LTS, Python 3.6.8, PyTorch 1.0.0, and CUDA 9.0.

Play with the Physics Engines

We provide three environments (1) Newton's Cradle, (2) Rope Manipulation, and (3) Box Pushing.

python physics_engine.py --env Cradle
python physics_engine.py --env Rope
python physics_engine.py --env Box

The visualizations will be stored in test/test_data_[env].

Evaluation

You can direct run the following command to use the pretrained checkpoint. Note that the provided checkpoints perform slightly better than what have been reported in the original paper, as we have performed more intensive hyperparameter tuning.

bash scripts/eval_Cradle.sh
bash scripts/eval_Rope.sh
bash scripts/eval_Box.sh

The resulting rollouts will be stored in dump_[env]/eval_[env]_pstep_[pstep]_*/, where ground truth is shown in transparent. Note that in env Box, this evaluation only shows the decoded results from the encoding space.

Model-Predictive Control

You can direct run the following command to use the pretrained checkpoint.

bash scripts/mpc_Box.sh
bash scripts/mpc_Rope.sh

The controlling result will be stored in dump_[env]/mpc_[env]_pstep_[pstep]_*/, where the target configuration is shown in transparent.

Training

You can use the following command to train from scratch. Note that if you are running script for the first time, it will start by generating training and validation data. You will need to change --gen_data to 0 if the data has already been generated.

bash scripts/train_Cradle.sh
bash scripts/train_Rope.sh
bash scripts/train_Box.sh

Citing PropNet

If you find this codebase useful in your research, please consider citing:

@inproceedings{li2019propagation,
    Title={Propagation Networks for Model-Based Control Under Partial Observation},
    Author={Li, Yunzhu and Wu, Jiajun and Zhu, Jun-Yan and Tenenbaum, Joshua B and Torralba, Antonio and Tedrake, Russ},
    Booktitle = {ICRA},
    Year = {2019}
}