Unsupervised Learning of Lagrangian Dynamics from Images for Prediction and Control
Authors: Yaofeng Desmond Zhong, Naomi Leonard
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We train our model on three systems: the Pendulum, the fully-actuated Cart Pole and the fully-actuated Acrobot. The training images are generated by Open AI Gym simulator [33]. The training setup is detailed in Supplementary Materials. As the mean square error in the image space is not a good metric of long term prediction accuracy [8], we report on the prediction image sequences of a previously unseen initial condition and highlight the interpretability of our model. Table 1: Average pixel MSE of different models. |
| Researcher Affiliation | Academia | Yaofeng Desmond Zhong Princeton University y.zhong@princeton.edu Naomi Ehrich Leonard Princeton University naomi@princeton.edu |
| Pseudocode | No | The paper describes the model architecture and training process in text and mathematical equations, but it does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | We implement this work with Py Torch [24] and refactor our code into Py Torch Lightning format [25], which makes our code easy to read and our results easy to reproduce. The code for all experiments is available at https://github.com/Desmond Zhong/Lagrangian_ca VAE. |
| Open Datasets | Yes | The training images are generated by Open AI Gym simulator [33]. |
| Dataset Splits | No | Table 1 mentions 'train | test' but the paper does not provide specific percentages or counts for training, validation, or test dataset splits in the main text. It states 'The training setup is detailed in Supplementary Materials', but these details are not present in the provided paper text. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU models, CPU types, or memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions that the work is implemented 'with Py Torch [24] and refactor our code into Py Torch Lightning format [25]', but it does not specify version numbers for these software dependencies. |
| Experiment Setup | Yes | Figure 3 shows the prediction sequences of images up to 48 time steps of the Pendulum and Cart Pole experiments with models trained with Tpred = 4. |