SoftZoo: A Soft Robot Co-design Benchmark For Locomotion In Diverse Environments

Authors: Tsun-Hsuan Wang, Pingchuan Ma, Andrew Everett Spielberg, Zhou Xian, Hao Zhang, Joshua B. Tenenbaum, Daniela Rus, Chuang Gan

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this work, we introduce Soft Zoo, a soft robot co-design platform for locomotion in diverse environments. ... We benchmark prevalent representations and co-design algorithms, and shed light on 1) the interplay between environment, morphology, and behavior 2) the importance of design space representations 3) the ambiguity in muscle formation and controller synthesis and 4) the value of differentiable physics. ... Algorithmic benchmarks for various representations and learning algorithms, laying a foundation for studying behavioral and morphological intelligence. Analysis of 1) the interplay between environment, morphology, and behavior 2) the importance of design space representations 3) the ambiguity in muscle formation and controller synthesis 4) the value of differentiable physics, with numerical comparisons of gradient-based and gradient-free design algorithms and intelligible examples of where gradient-based co-design fails.
Researcher Affiliation Collaboration Tsun-Hsuan Wang1, , Pingchuan Ma1, Andrew Spielberg1,4, Zhou Xian5, Hao Zhang3, Joshua B. Tenenbaum1, Daniela Rus1, Chuang Gan2,3 1MIT CSAIL, 2MIT-IBM Watson AI Lab, 3UMass Amherst, 4Harvard, 5CMU
Pseudocode No The paper describes algorithms and methods using prose and mathematical equations but does not include any explicitly labeled "Pseudocode" or "Algorithm" blocks.
Open Source Code Yes We will release the code-level implementation of all materials for reproducibility.
Open Datasets No The paper discusses generating environments and using
Dataset Splits No The paper does not provide specific train/validation/test dataset splits. It describes various optimization processes (e.g., control optimization, design optimization) and the number of training iterations, but these are distinct from dataset partitioning for reproduction.
Hardware Specification No The paper mentions that the Taichi library "compiles physical simulation code (and its reverse-mode autodifferentiated gradients) for efficient parallelization on GPU hardware." However, it does not specify any particular GPU models (e.g., NVIDIA A100, RTX 2080 Ti), CPU models, or other specific hardware configurations used for the experiments.
Software Dependencies No The paper mentions using specific software libraries and methods such as "Taichi library (Hu et al., 2019b)", "Proximal Policy Optimization (PPO) (Schulman et al., 2017)", and "Adam as the optimizer". However, it does not provide specific version numbers for these software dependencies, which are necessary for reproducible setup.
Experiment Setup Yes For the large-scale benchmark with biologically-inspired design (Table 1), we use learning rate 0.1 and training iterations 30. For all control-only and design-only optimization, we use learning rate 0.01 and training iterations 100. For co-design, we use learning rate 0.01 for both control and design with training iterations 250. We use Adam as the optimizer. RL is only used in control optimization. We use Proximal Policy Optimization (PPO) (Schulman et al., 2017) with the following hyperparameters: number of timesteps 10^5, buffer size 2048, batch size 32, GAE coefficient 0.95, discounting factor 0.98, number of epochs 20, entropy coefficient 0.001, learning rate 0.0001, clip range 0.2. ... We run ES for 100 generations for the co-design baseline.