Differentiable Simulation of Soft Multi-body Systems
Authors: Yiling Qiao, Junbang Liang, Vladlen Koltun, Ming Lin
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experiments demonstrate that our designs make soft body simulation more stable and realistic compared to other frameworks. Our method accelerates the solution of system identification problems by more than an order of magnitude, and enables efficient gradient-based learning of motion control with soft robots. |
| Researcher Affiliation | Collaboration | Yi-Ling Qiao University of Maryland, College Park Junbang Liang University of Maryland, College Park Vladlen Koltun Intel Labs Ming C. Lin University of Maryland, College Park |
| Pseudocode | Yes | Algorithm 1 Soft body simulation using Projective Dynamics; Algorithm 2 Matrix Assembly for the Articulated System |
| Open Source Code | Yes | Code is available on our project page: https://github.com/YilingQiao/diff_fem |
| Open Datasets | No | The paper describes simulated scenarios (Baymax model, bridges, pneumatic gripper, octopus, fish) but does not provide concrete access information (link, DOI, specific citation with authors/year) for any publicly available dataset used for training. |
| Dataset Splits | No | The paper does not provide specific details on dataset splits (e.g., percentages or counts) for training, validation, or testing. |
| Hardware Specification | Yes | We run our experiments on two desktops, one with an Intel Xeon W-2123 CPU @ 3.6GHz and the other with an Intel i9-10980XE @ 3.0GHz, respectively. |
| Software Dependencies | No | Our simulator is written in C++, the learning algorithms are implemented in PyTorch [60], and Pybind [34] is used as the interface... Cpp AD [3] is used to differentiate the simulation here. The paper mentions these software components but does not provide specific version numbers for them as used in their implementation. |
| Experiment Setup | Yes | Each experiment is repeated 5 times with different random seeds. ... The pneumatic gripper in Figure 4(a) has 56 pneumatic cells in four arms... We control the pneumatic activation as well as the movement of the drone to move the ball from the start (0, 0, 0) to our target (0, 0.3, 0) in 50 steps. The loss is the distance from the actual position to the target position. ... The length of the simulation is 400 steps, and the control input in each step is 64-dimensional. In total, there are 64 * 400 = 25600 variables to optimize. ... The fish... starts at (0, 0, 0) in step 1 and the target location in step 100 is (0, 0, 0.15). The objective function is the distance from the actual location to the target location. For each step, there will be a torque vector of size 5 that represents the joint actuation level. In total, the optimization variable has 500 dimensions. |