Visual Grounding of Learned Physical Models

Yunzhu Li, Toru Lin, Kexin Yi, Daniel M. Bear, Daniel L. K. Yamins, Jiajun Wu, Joshua B. Tenenbaum, and Antonio Torralba

ICML 2020 [website] [paper] [video]

Evaluate the trained model on the test data

Create a folder named data/ in the project folder and download the test data from the following links. Unzip and put them in data/.

MassRope [DropBox] (2.8 GB)
RigidFall [DropBox] (7.35 GB)

1. Dynamics Prior

Type the following commands to evaluate the dynamics prior. You can also take a look at [VGPL-Dynamics-Prior], which we prepared as a stand-alone module for dynamics prediction.

bash scripts/dynamics/eval_MassRope_dy.sh
bash scripts/dynamics/eval_RigidFall_dy.sh

You will be able to generate videos like the following

2. Visual Prior

Please check [VGPL-Visual-Prior] for details of the visual prior, which estimates the particle positions and groupings from the visual observations.

For the ease of evaluation, we have included the visual prior's prediction results in the test data you just downloaded: data_[ENV]/perception/l2. You can proceed to run the commands in the following sections to see the evaluation results of the inference module.

3. Parameter Estimation

Type the following command to evaluate the model's performance on parameter estimation on 50 testing examples.

bash scripts/parameter/eval_MassRope_param.sh
bash scripts/parameter/eval_RigidFall_param.sh

MassRope: Average error ratio: 2.8812% (std: 1.2736)
RigidFall: Average error ratio: 3.7455% (std: 2.6649)

4. Position Refinement and Rigidness estimation

Type the following command to evaluate the model's performance on position refinement and rigidness estimation on 50 testing examples.

bash scripts/position/eval_MassRope_pos.sh
bash scripts/position/eval_RigidFall_pos.sh

Position Mean Squared Error (scaled by 1e4):

MassRope: Before refinement 1.9584, After refinement 0.4782
RigidFall: Before refinement 1.9700, After refinement 1.4500

5. Forward prediction using the inference results

Type the following command to evaluate the model's performance on forward prediction using the inference results.

bash scripts/forward/eval_MassRope_fwd.sh
bash scripts/forward/eval_RigidFall_fwd.sh

The generated videos are stored in dump/dump_{ENV}/eval_forward*. The following are two demos, where we show the original RGB image, the prediction, and the ground truth from left to right.

Train the model

For training the dynamics prior, please refer to [VGPL-Dynamics-Prior], which we prepared as a stand-alone module for dynamics prediction.

For training the visual prior, please refer to [VGPL-Visual-Prior].

For training the dynamics-guided inference module, we have provided the following training scripts

scripts/parameter/train_MassRope_param.sh
scripts/position/train_MassRope_pos.sh

scripts/parameter/train_RigidFall_param.sh
scripts/position/train_RigidFall_pos.sh

To obtain the training data, which is about nine times the size of the test data, please contact me through liyunzhu@mit.edu.

Citing VGPL

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

@inproceedings{li2020visual,
    Title={Visual Grounding of Learned Physical Models},
    Author={Li, Yunzhu and Lin, Toru and Yi, Kexin and Bear, Daniel and Yamins, Daniel L.K. and Wu, Jiajun and Tenenbaum, Joshua B. and Torralba, Antonio},
    Booktitle={ICML},
    Year={2020}
}

@inproceedings{li2019learning,
    Title={Learning Particle Dynamics for Manipulating Rigid Bodies, Deformable Objects, and Fluids},
    Author={Li, Yunzhu and Wu, Jiajun and Tedrake, Russ and Tenenbaum, Joshua B and Torralba, Antonio},
    Booktitle={ICLR},
    Year={2019}
}

YunzhuLi / VGPL

readme