Visual Grounding of Learned Physical Models
Authors: Yunzhu Li, Toru Lin, Kexin Yi, Daniel Bear, Daniel Yamins, Jiajun Wu, Joshua Tenenbaum, Antonio Torralba
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our model in environments involving interactions between rigid objects, elastic materials, and fluids. Experiments demonstrate that our model, within a few observation steps, is able to refine the particle positions proposed by the visual prior, accurately predict the rigidness the objects, and infer the physical parameters, which enables quick adaptation to new scenarios with unknown physical properties and making predictions into the future. and 4. Experiments We study our framework under three environments that incorporate different types of objects and facilitate rich interactions. In this section, we show results and present ablation studies on various inference and prediction tasks. |
| Researcher Affiliation | Academia | 1MIT CSAIL 2Harvard University 3Wu Tsai Neurosciences Institute and Department of Psychology, Stanford University 4Department of Computer Science, Stanford University 5MIT BCS, CBMM, CSAIL. |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions 'Please check our project page for video demonstrations.' but does not provide an explicit statement or link for the release of their source code. |
| Open Datasets | No | We use NVIDIA Fle X (Macklin et al., 2014), a particle-based physics engine to generate all data for training and testing. The paper indicates data was generated but does not provide concrete access information for a publicly available or open dataset. |
| Dataset Splits | No | For all three environments, we use 90% of the data for training and 10% for testing. No explicit mention of a separate validation split. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running its experiments, only mentioning general implementation details. |
| Software Dependencies | No | All models are implemented in Py Torch (Paszke et al., 2019) and trained with the Adam optimizer (Kingma & Ba, 2015). While software is mentioned, specific version numbers for PyTorch or other dependencies are not provided. |
| Experiment Setup | Yes | We use a batch size of 50 and a learning rate of 10 4 to train the model for 2700 iterations on all environments. The particle positions are normalized. The sequence length of input and output data per forward pass is set to be 4. and We use a batch size of 4 and a learning rate of 10 5. The model observes 4 past time steps and predicts 1 time step into the future. and Our inference module is trained for 2 epochs on each environment, using a batch size of 2 and a learning rate of 10 5. Length of the input and output sequences is set to be T = 10. |