Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Gaussian-Augmented Physics Simulation and System Identification with Complex Colliders
Authors: Federico Vasile, Ri-Zhao Qiu, Lorenzo Natale, Xiaolong Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments Physical Parameters. Our analysis focuses on material types that undergo noticeable deformation upon collision with rigid bodies, specifically Newtonian fluids, non-Newtonian fluids, and granular media. For Newtonian fluids, we estimate fluid viscosity (µ) and bulk modulus (κ). For non Newtonian fluids, we recover shear modulus (µ), bulk modulus (κ), yield stress (τY ), and plasticity viscosity (η). For granular media, we estimate the friction angle (θfric). For a comprehensive overview of physical parameters and constitutive models, we refer to [4]. Dataset. We follow the protocol in [4] and generate ground-truth simulation rollouts using the cross-shaped object from [4] as our continuum material, with the physical parameter values provided in Appendix E. The object undergoes free fall and collides with a static rigid body (Box, Bunny, or Armadillo) having sticky surfaces. For each rollout, we collect both the 3D particle trajectories (i.e., point clouds over time) of the continuum material and the corresponding rendered frames (see Fig. 3), captured from 11 cameras placed around the scene. These data are used in Sec. 5.1 and Sec. 5.2 for experiments on system identification. Following the protocol in [4], we generate 10 rollouts per collider for each material type, except for granular media, where 5 videos are generated, resulting in a total of 75 rollouts. Each rollout is 16 timesteps long. Training and Evaluation. Our training setup also adheres to [4], including the use of the Adam optimizer [35] and the initial guesses for the physical parameters. Unlike [4], we do not optimize the initial velocity vector during the free-fall phase; instead, we use the ground truth values, as our focus is on system identification during collisions. Finally, we report the results of physical parameters estimation using the mean and standard deviation of absolute errors, scaled by a factor of 100. |
| Researcher Affiliation | Academia | Federico Vasile1 Ri-Zhao Qiu2 Lorenzo Natale1 Xiaolong Wang2 1Istituto Italiano di Tecnologia, 2UC San Diego |
| Pseudocode | No | The paper describes methods and procedures using prose and mathematical equations but does not present them in a structured pseudocode or algorithm block format. |
| Open Source Code | Yes | https://as-diffmpm.github.io Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: Instructions on how to download the data and run scripts are not included in the paper, however, we will rely on a project page (linked to our paper) to provide instructions for reproducibility. We well release ready-to-use data and scripts (e.g., no data pre-processing required). Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? Answer: [Yes] Justification: We did our best to include all the details for reproducibility. Moreover, we will release the code of our framework. |
| Open Datasets | No | Dataset. We follow the protocol in [4] and generate ground-truth simulation rollouts using the cross-shaped object from [4] as our continuum material, with the physical parameter values provided in Appendix E. The object undergoes free fall and collides with a static rigid body (Box, Bunny, or Armadillo) having sticky surfaces. For each rollout, we collect both the 3D particle trajectories (i.e., point clouds over time) of the continuum material and the corresponding rendered frames (see Fig. 3), captured from 11 cameras placed around the scene. These data are used in Sec. 5.1 and Sec. 5.2 for experiments on system identification. Following the protocol in [4], we generate 10 rollouts per collider for each material type, except for granular media, where 5 videos are generated, resulting in a total of 75 rollouts. Each rollout is 16 timesteps long. Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: Instructions on how to download the data and run scripts are not included in the paper, however, we will rely on a project page (linked to our paper) to provide instructions for reproducibility. We well release ready-to-use data and scripts (e.g., no data pre-processing required). |
| Dataset Splits | No | Dataset. We follow the protocol in [4] and generate ground-truth simulation rollouts using the cross-shaped object from [4] as our continuum material, with the physical parameter values provided in Appendix E. The object undergoes free fall and collides with a static rigid body (Box, Bunny, or Armadillo) having sticky surfaces. For each rollout, we collect both the 3D particle trajectories (i.e., point clouds over time) of the continuum material and the corresponding rendered frames (see Fig. 3), captured from 11 cameras placed around the scene. These data are used in Sec. 5.1 and Sec. 5.2 for experiments on system identification. Following the protocol in [4], we generate 10 rollouts per collider for each material type, except for granular media, where 5 videos are generated, resulting in a total of 75 rollouts. Each rollout is 16 timesteps long. |
| Hardware Specification | Yes | F Implementation Details We implemented our framework using Python and Taichi for differentiable programming and parallel computation. AS-Diff MPM is built upon the open source Diff MPM implementation in [4] and subsequent works [7, 8]. We run the experiments on NVIDIA RTX 3080 and 4090 graphics cards. |
| Software Dependencies | No | F Implementation Details We implemented our framework using Python and Taichi for differentiable programming and parallel computation. AS-Diff MPM is built upon the open source Diff MPM implementation in [4] and subsequent works [7, 8]. We run the experiments on NVIDIA RTX 3080 and 4090 graphics cards. |
| Experiment Setup | Yes | Training and Evaluation. Our training setup also adheres to [4], including the use of the Adam optimizer [35] and the initial guesses for the physical parameters. Unlike [4], we do not optimize the initial velocity vector during the free-fall phase; instead, we use the ground truth values, as our focus is on system identification during collisions. Finally, we report the results of physical parameters estimation using the mean and standard deviation of absolute errors, scaled by a factor of 100. Table 13: Values of physical parameters for ground-truth simulation rollouts. Newtonian (Initial Guess: µ = 10, κ = 10^4) Non-Newtonian (Initial Guess: µ = 100, κ = 10^5, τY = 10, η = 1) Granular (Initial Guess: θfric = 10) |