gradSim: Differentiable simulation for system identification and visuomotor control
Authors: J. Krishna Murthy, Miles Macklin, Florian Golemo, Vikram Voleti, Linda Petrini, Martin Weiss, Breandan Considine, Jérôme Parent-Lévesque, Kevin Xie, Kenny Erleben, Liam Paull, Florian Shkurti, Derek Nowrouzezahrai, Sanja Fidler
ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate Sim s effectiveness on parameter identification tasks for rigid, deformable and thinshell bodies, and demonstrate performance that is competitive, or in some cases superior, to current physics-only differentiable simulators. Additionally, we demonstrate the effectiveness of the gradients provided by Sim on challenging visuomotor control tasks involving deformable solids and cloth. We conducted multiple experiments to test the efficacy of Sim on physical parameter identification from video and visuomotor control |
| Researcher Affiliation | Collaboration | Krishna Murthy Jatavallabhula 1,3,4, Miles Macklin 2, Florian Golemo1,3, Vikram Voleti3,4, Linda Petrini3, Martin Weiss3,4, Breandan Considine3,5, Jérôme Parent-Lévesque3,5, Kevin Xie2,6,7, Kenny Erleben8, Liam Paull1,3,4, Florian Shkurti6,7, Derek Nowrouzezahrai3,5, and Sanja Fidler2,6,7 1Montreal Robotics and Embodied AI Lab, 2NVIDIA, 3Mila, 4Université de Montréal, 5Mc Gill, 6University of Toronto, 7Vector Institute, 8University of Copenhagen |
| Pseudocode | Yes | Listing 1: Particle Integration Kernel |
| Open Source Code | Yes | https://gradsim.github.io |
| Open Datasets | Yes | We curate a dataset of 10000 simulated videos generated from variations of 14 objects, comprising primitive shapes such as boxes, cones, cylinders, as well as non-convex shapes from Shape Net (Chang et al., 2015) and Dex Net (Mahler et al., 2017). |
| Dataset Splits | No | Across all rigid-body experiments, we use 800 objects for training and 200 objects for testing. |
| Hardware Specification | Yes | We find that, on a laptop with an Intel i7 processor and a Ge Force GTX 1060 GPU, parameter estimation experiments for rigid/nonrigid bodies can be run in under 5-20 minutes per object on CPU and in under 1 minute on the GPU. |
| Software Dependencies | No | The paper mentions 'Py Torch framework' and cites the PyTorch paper from 2019, but does not provide a specific version number (e.g., PyTorch 1.x). It also mentions 'CUDA Kirk et al. (2007)' and 'Py Bullet Coumans & Bai (2016 2019)' without specific version numbers. |
| Experiment Setup | Yes | Inference with Sim is done by guessing an initial mass (uniformly random in the range [2, 12]kg/m3), unrolling a differentiable simulation using this guess, comparing the rendered out video with the true video (pixelwise mean-squared error MSE), and performing gradient descent updates. ... We use Efficient Net (B0) ... and resize input frames to 64 64. ...We train the model to minimize the mean-squared error between the estimated and the true parameters, and use the Adam ... optimizer with learning rate of 0.0001. Each model was trained for 100 epochs on a V 100 GPU. The hyperparameters used in this baseline can be found in Table 5. |