Learning rigid dynamics with face interaction graph networks

Authors: Kelsey R Allen, Yulia Rubanova, Tatiana Lopez-Guevara, William F Whitney, Alvaro Sanchez-Gonzalez, Peter Battaglia, Tobias Pfaff

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Compared to learned node-and particle-based methods, FIGNet is around 4x more accurate in simulating complex shape interactions, while also 8x more computationally efficient on sparse, rigid meshes. We conduct a series of ablations and experiments which showcase how face-to-face collision representations dramatically improve rigid body dynamics prediction.
Researcher Affiliation Industry Kelsey R. Allen , Yulia Rubanova , Tatiana Lopez-Guevara, William Whitney, Alvaro Sanchez-Gonzalez, Peter Battaglia, Tobias Pfaff Deep Mind, London, UK
Pseudocode No The paper describes the model architecture and algorithms using prose and mathematical equations, but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code No The paper does not contain an explicit statement that the source code for FIGNet is being released or a direct link to its repository.
Open Datasets Yes The Kubric dataset (Greff et al., 2022) consists of the rigid-body simulations of diverse 3D objects tossed simultaneously onto a large plane. For these experiments, we use the MIT Pushing Dataset ((Yu et al., 2016); Figure 7a)...
Dataset Splits Yes For each of MOVi-A, MOVi-B and Movi C datasets, we use 1500 trajectories for training, 100 for validation and 100 for testing.
Hardware Specification Yes All models are trained to 1M steps with a batch size of 128 across 8 TPU devices.
Software Dependencies No The paper mentions using components like Adam optimizer and Layer Norm, but does not provide specific version numbers for software libraries or frameworks like PyTorch or TensorFlow.
Experiment Setup Yes All models are trained to 1M steps with a batch size of 128 across 8 TPU devices. We use Adam optimizer, and an an exponential learning rate decay from 1e-3 to 1e-4. To stabilize long rollouts, we add random-walk noise to the positions during training (Sanchez-Gonzalez et al., 2020; Pfaff et al., 2021; Allen et al., 2022).