Learning Deformable Body Interactions With Adaptive Spatial Tokenization
Authors: Hao Wang, Yu Liu, Daniel Biggs, Haoru Wang, Jiandong Yu, Ping Huang
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our method significantly outperforms state-of-the-art methods in modeling deformable body interactions. Notably, it remains effective on large-scale simulations with meshes exceeding 100,000 nodes, where existing methods are hindered by computational limitations. Additionally, we contribute a novel large-scale dataset encompassing a wide range of deformable body interactions to support future research in this area. |
| Researcher Affiliation | Industry | Hao Wang Yu Liu Daniel Biggs Apple Apple Apple EMAIL EMAIL EMAIL Haoru Wang Jiandong Yu Ping Huang Apple Apple Apple EMAIL EMAIL EMAIL |
| Pseudocode | Yes | A.6 Farthest Point Sampling Implementation We outline the FPS algorithm in Algorithm 1, while noting that readers may directly use existing open-source implementations. Algorithm 1 Farthest Point Sampling (FPS) Require: Point set P = {p1, . . . , p N}, sample size K Ensure: Sample indices S of size K 1: Initialize S {}. 2: Randomly choose initial index i0. 3: S S {i0}. 4: for each i {1, . . . , N} \ S do 5: d[i] dist(pi, pi0) 6: end for 7: for t 2 to K do 8: j arg max i S d[i] 9: S S {j} 10: for each i S do 11: d[i] min d[i], dist(pi, pj) 12: end for 13: end for 14: return S |
| Open Source Code | No | The paper does not provide an explicit statement or link for the open-sourcing of its own methodology's code. It mentions using existing open-source implementations for Farthest Point Sampling but not for the core AST method. |
| Open Datasets | Yes | Datasets We adopt two representative public datasets from Graph Mesh Nets (Pfaff et al., 2020) that involve object interactions: 1) Deforming Plate: A deformable object is compressed by a rigid body, with ~1.3k mesh/4k element nodes per mesh; 2) Sphere Simple: A piece of cloth interacts with a kinematic sphere, with ~2k mesh/4k element nodes per mesh. To further validate our method on large-scale physical simulation tasks an area where existing literature is limited we introduce two new datasets: 3)ABCD: ABCD stands for A Big CAD Deformation, where two deformable objects squish each other, with ~4k mesh/12k element nodes per mesh; 4) ABCD-XL: follows the same setup as the Abcd dataset, except it uses significantly denser meshes, with ~100k mesh/300k element nodes per mesh. ... Additionally, we contribute a novel large-scale dataset encompassing a wide range of deformable body interactions to support future research in this area. |
| Dataset Splits | Yes | Table 1: Details of the datasets used in this work. ... #Samples indicates the number of simulations generated for training, validation, and testing, respectively. Deforming Plate ... 1000 : 100 : 100 Sphere Simple ... 1000 : 100 : 100 ABCD ... 6000 : 300 : 300 ABCD-XL ... 1000 : 100 : 100 |
| Hardware Specification | Yes | All experiments on Deforming Plate, Sphere Simple, and Abcd are conducted on a single machine equipped with 4 V100 GPUs. For the large-scale ABCD-XL dataset, experiments are run on a machine with 8 V100 GPUs. Other training details can be found in Section A.1 and Section A.2. ... All experiments are conducted on a machine equipped with four V100-32GB GPUs, unless otherwise specified. |
| Software Dependencies | No | The paper mentions various architectural components and techniques such as MLPs, ReLU activations, Layer Norm, Random Weight Factorization, GEGLU activation, and Dropout, but it does not specify any particular software libraries or their version numbers (e.g., PyTorch version, CUDA version). |
| Experiment Setup | Yes | A.2 Training Settings For all datasets, we adopt a pairwise training strategy where a graph is randomly selected from a sequence as the input, and its subsequent graph is used as the target. We follow the same training noise strategy as proposed in Pfaff et al. (2020). The noise scale is set to 0.003 for both the Abcd and Abcd-XL datasets. The input and target features are all normalized to zero mean and unit variance based on the statistics of the training set. ... Batch Size We increased the training batch size from 1 or 2 to 48 (12 per GPU on a 4-GPU node) for MGN and BSMS which showed a much faster training procedure. HCMT inherently not applicable on batched graphs, so we kept a batch size with 1 per GPU. Learning Rate We adopt square root scaling for the learning rate with respect to batch size. Starting with a base learning rate of 0.0001 for a batch size of 2, the final learning rate LR for a batch size of 48 is computed as 0.0001 p 48/2 0.00049. The learning rate is linearly warmed up from 0.0001LR to LR over the first 2000 steps, followed by cosine decay to zero at the 101st epoch (training stops at the 100th epoch). ... Training Iterations We extend the training iterations from 5M steps (approximately 25 epochs) to 100 epochs. Loss We use mean squared error (MSE) loss across all experiments. For the Deforming Plate and Abcd-XL datasets, we adopt multi-head outputs to jointly predict displacement and stress, assigning loss weights of 1 and 0.01, respectively. ... A.5 Model Details MLP The MLPs used in input, message-passing and output layers are two-layer MLPs with Re LU activations with output size of 128. The hidden layer size of message-passing and input MLPs are 128, while the output MLPs are 32. ... Transformer We use attention layers with 8 heads, each with a feature dimension of 64. The query, key, and value projections are implemented using bias-free linear layers, while the output projection includes a bias term. We adopt a pre-norm setup, applying Layer Norm before both the attention and FFN layers. The FFN follows the GEGLU formulation as described in Shazeer (2020), and has a hidden dimension of 512. Dropout with a rate of 0.1 is applied within both the attention and FFN modules. Dataset Lcell locnn dtoken LSA Sphere Simple 5 0 256 12 Deforming Plate 5 0 256 12 Abcd 8 0 512 12 Abcd-XL 12 4 512 12 Figure 14: Model hyperparameter. |