Learning Articulated Rigid Body Dynamics with Lagrangian Graph Neural Network

Authors: Ravinder Bhattoo, Sayan Ranu, N M Anoop Krishnan

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we evaluate the ability of LGNN to learn rigid body dynamics. In addition, we evaluate the ability of LGNN to generalize to larger unseen system sizes, complex topology, and realistic structures such as tensegrity.
Researcher Affiliation Academia Ravinder Bhattoo Department of Civil Engineering Indian Institute of Technology Delhi cez177518@iitd.ac.in Sayan Ranu Department of Computer Science and Engineering Yardi School of Arti cial Intelligence Indian Institute of Technology Delhi sayanranu@iitd.ac.in N. M. Anoop Krishnan Department of Civil Engineering Yardi School of Arti cial Intelligence Indian Institute of Technology Delhi krishnan@iitd.ac.in
Pseudocode No The paper describes the model architecture and equations but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes The code is available at https://github.com/M3RG-IITD/rigid_body_dynamics_graph
Open Datasets No The paper generates its own dataset for experiments and does not explicitly state that this generated dataset is publicly available or open for access. While the code to generate the data is available, the dataset itself is not provided with concrete access information as a separate entity.
Dataset Splits Yes This dataset is divided randomly in 75:25 ratio as training and validation set.
Hardware Specification Yes Hardware: Memory: 16Gi B System memory, Processor: Intel(R) Core(TM) i7-10750H CPU @ 2.60GHz
Software Dependencies Yes Software packages: numpy-1.20.3, jax-0.2.24, jax-md-0.1.20, jaxlib-0.1.73, jraph-0.0.1.dev0
Experiment Setup Yes For the graph architectures, namely, LGNN and GNS, all the neural networks are modeled as one hidden layer MLPs with varying number of hidden units. For all the MLPs, a square-plus activation function is used due to its double differentiability. In contrast to the earlier approaches, here, the training is not performed on trajectories. Rather, it is performed on 10000 data points generated from 100 trajectories for all the models. This dataset is divided randomly in 75:25 ratio as training and validation set. The model performance is evaluated on a forward trajectory, a task it was not explicitly trained for, of 1s.