FluidLab: A Differentiable Environment for Benchmarking Complex Fluid Manipulation
Authors: Zhou Xian, Bo Zhu, Zhenjia Xu, Hsiao-Yu Tung, Antonio Torralba, Katerina Fragkiadaki, Chuang Gan
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate model-free RL algorithms, sampling-based optimization methods as well as trajectory optimization using differentiable physics coupled with our proposed optimization techniques in Fluid Lab. We discuss our optimization techniques and result findings in this Section. For detailed loss and reward design, please refer to Appendix B.2. We evaluate our proposed optimization schemes coupled with differentiable physics (DP), modelfree RL algorithms including Soft Actor-Critic (SAC) (Haarnoja et al., 2018) and Proximal Policy Optimization (PPO) (Schulman et al., 2017), CMA-ES (Hansen & Ostermeier, 2001), a sampling-based trajectory optimization method, as well as PODS (Mora et al., 2021) an method combines RL and differentiable simulation (see Appendix C for a discussion). We use the open source implementation (Raffin et al., 2021) for PPO and SAC, and pycma (Hansen et al., 2019) for CMA-ES. |
| Researcher Affiliation | Collaboration | Zhou Xian CMU zhouxian@cmu.edu Bo Zhu Dartmouth College bo.zhu@dartmouth.edu Zhenjia Xu Columbia University xuzhenjia@cs.columbia.edu Hsiao-Yu Tung MIT hytung@mit.edu Antonio Torralba MIT torralba@mit.edu Katerina Fragkiadaki CMU katef@cs.cmu.edu Chuang Gan MIT-IBM Watson AI Lab ganchuang@csail.cmu.edu |
| Pseudocode | No | The paper describes methods and techniques but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | Fluid Lab is publicly available at: https://fluidlab2023.github.io. |
| Open Datasets | No | While the Fluid Lab environment itself is publicly available, the paper does not explicitly state that a pre-collected 'dataset' used for training or evaluation is provided via a link or citation. The experiments are conducted within the publicly available environment. |
| Dataset Splits | No | The paper describes evaluation of RL algorithms and trajectory optimization methods within the Fluid Lab environment, but does not explicitly specify fixed training, validation, and test dataset splits for reproducibility in the context of static data. |
| Hardware Specification | Yes | All our proposed tasks contain around 100,000 particles and runs in real time (60 FPS where each frame is a simulation step) on a desktop computer equipped with an Nvidia RTX 3090 GPU and an Intel i7-8700K CPU, and the typical GPU usage is under 50%. |
| Software Dependencies | No | The paper mentions software like Taichi, Python, Open AI Gym API, PPO, SAC, and pycma, but it does not provide specific version numbers for these software components. It only cites the original papers or projects, not the exact versions used in their experimental setup. |
| Experiment Setup | Yes | All our tasks use a simulation step of 2e-3 seconds, each containing 10 substeps of 1e-4 seconds for ensuring simulation stability. All the tasks use an MPM grid of size 64^3, with around 100,000 particles, and a 128^3 grid for gas simulation. We discuss our optimization techniques and result findings in this Section. |