Structured Object-Aware Physics Prediction for Video Modeling and Planning
Authors: Jannik Kossen, Karl Stelzner, Marcel Hussing, Claas Voelcker, Kristian Kersting
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | STOVE predicts videos with convincing physical behavior over thousands of timesteps, outperforms previous unsupervised models, and even approaches the performance of supervised baselines. We further demonstrate the strength of our model as a simulator for sample efficient model-based control in a task with heavily interacting objects. |
| Researcher Affiliation | Academia | 1Department of Physics and Astronomy, Heidelberg University 1kossen@stud.uni-heidelberg.de 2,3Department of Computer Science, TU Darmstadt 2{stelzner,kersting}@cs.tu-darmstadt.de 3{marcel.hussing,c.voelcker}@stud.tu-darmstadt.de |
| Pseudocode | No | The paper describes network architectures and processes but does not include formal pseudocode or algorithm blocks. |
| Open Source Code | Yes | With this paper, we also release a Py Torch implementation of STOVE.1 The code can be found in the Git Hub repository github.com/jlko/STOVE. |
| Open Datasets | No | For the billiards and gravitational data, 1000 sequences of length 100 were generated for training. From these, subsequences of lengths 8 were sampled and used to optimize the ELBO. A test dataset of 300 sequences of length 100 was also generated and used for all evaluations. The paper generates its own dataset but does not provide access information for it. |
| Dataset Splits | No | The paper mentions generating data for training and testing but does not explicitly describe a separate validation set or its split. |
| Hardware Specification | Yes | When trained using a single GTX 1080 Ti, STOVE converges after about 20 hours. |
| Software Dependencies | No | The paper states it releases a PyTorch implementation but does not specify the version of PyTorch or other software dependencies with version numbers used for their own work. |
| Experiment Setup | Yes | Our model was trained using the Adam optimizer (Kingma & Ba, 2015), with a learning rate of 2 10 3 exp( 40 10 3 step) for a total of 83 000 steps with a batch size of 256. The training uses an Adam optimizer with a learning rate of 2 10 4 and and ϵ value of 1 10 5. The clipping parameter of PPO is set to 1 10 1. We update the network for 4 epochs in each batch using 32 mini-batches of the sampled data. The value loss is weighted at 5 10 1 and the entropy coefficient is set to 1 10 2. |