See arxiv.org/1910.02425 or openreview.net/forum?id=B1e-kxSKDH for further information.
| Real | Ours | VRNN | SQAIR | DDPAE | Linear | Supervised |
|---|
| Real | Ours | VRNN | SQAIR | DDPAE | Linear | Supervised |
|---|
The above depicts the reconstruction and prediction errors of the various models. The models are given 8 frames of video as input, which they reconstruct. Conditioned on this input, all models predict the following 92 frames. Only STOVE manages to generate visually convincing physically behavior over longer timeframes. 10 different sequences of length 100 are shown.
| MCTS on STOVE | MCTS on Real Env. | PPO on Env. States | PPO on Env. Images |
|---|
The above shows the performance of the compared models in the interactive environments. The agent controls the red ball and negative reward is given whenever the red ball collides with any other ball. STOVE is used as a world model, predicting future states, frames, and rewards. MCTS can then be used on STOVE for model-based control. We compare to MCTS on the real environment states, as well as PPO on the environment states and raw images. Again, 10 different sequences of length 100 are shown.
Humans can easily predict future outcomes even in settings with complicated interactions between objects. For computers, however, learning models of interactions from videos in an unsupervised fashion is hard. In this paper, we demonstrate that structure and compositionality are key to solving this problem. Specifically, we develop a novel video prediction model from two distinct components: a scene model and a physics model. Both models are compositional and exploit repeating elements wherever possible. We impose a highly structured bottleneck between model components to allow for fast learning and clearly interpretable functionality, without losing any generality or performance. This fully compositional approach yields a strong video prediction model, which clearly outperforms relevant baselines. We produce realistic looking physical behaviour over a possibly infinite time frame and perform competitively even compared to a supervised approach. Finally, we demonstrate the strength of our model as a simulator for sample efficient model-based reinforcement learning in tasks with heavily interacting objects.
Please cite our work here and at arxiv.org/1910.02425 as
@inproceedings{kossen2020structured,
title={Structured Object-Aware Physics Prediction for Video Modeling and Planning},
author={Kossen, Jannik and Stelzner, Karl and Hussing, Marcel and Voelcker, Claas and Kersting, Kristian},
booktitle={Proceedings of the International Conference on Learning Representations},
year={2020}
}
Run run_scripts.py --create-data to generate billiards and gravity data.
Also random data collected in the RL setting to train the action-
conditioned world model is generated.
Run bash run_files/run_models.sh to train the model on billiards and gravity data.
Also an actioned-conditioned world model is trained on the avoidance task.
python run_scripts.py --interactive allows you to either play in a live
environment or a model simulation of the environment.
If you have any questions or problems regarding the code or paper do not hesitate to contact us.