SOFTZOO: A SOFT ROBOT CO-DESIGN BENCHMARK FOR LOCOMOTION IN DIVERSE ENVIRONMENTS

Abstract

While significant research progress has been made in robot learning for control, unique challenges arise when simultaneously co-optimizing morphology. Existing work has typically been tailored for particular environments or representations. In order to more fully understand inherent design and performance tradeoffs and accelerate the development of new breeds of soft robots, a comprehensive virtual platform -with well-established tasks, environments, and evaluation metricsis needed. In this work, we introduce SoftZoo, a soft robot co-design platform for locomotion in diverse environments. SoftZoo supports an extensive, naturallyinspired material set, including the ability to simulate environments such as flat ground, desert, wetland, clay, ice, snow, shallow water, and ocean. Further, it provides a variety of tasks relevant for soft robotics, including fast locomotion, agile turning, and path following, as well as differentiable design representations for morphology and control. Combined, these elements form a feature-rich platform for analysis and development of soft robot co-design algorithms. 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. We envision that SoftZoo will serve as a standard platform and template an approach toward the development of novel representations and algorithms for co-designing soft robots' behavioral and morphological intelligence. Demos are available on our project page 1 .

1. INTRODUCTION

The natural world demonstrates morphological and behavioral complexity to a degree unexplored in soft robotics. A jellyfish's gently undulating geometry allows it to efficiently travel across large bodies of water; an ostrich's spring-like feet allow for fast, agile motion over widely varying topography; a chameleon's feet allows for dexterous climbing up trees and across branches. Beyond their comparative lack of diversity, soft robots' designs are rarely computationally optimized in silico for the environments in which they are to be deployed. The degree of morphological intelligence observed in the natural world would be similarly advantageous in artificial life. In this paper, we present SoftZoo, a framework for exploring and benchmarking algorithms for codesigning soft robots in behavior and morphology, with emphasis on locomotion tasks. Unlike pure control or physical design optimization, co-design algorithms co-optimize over a robot's brain and body simultaneously, finding more efficacious solutions that exploit their rich interplay (Ma et al., 2021; Spielberg et al., 2021; Bhatia et al., 2021) . We have seen examples of integrated morphology and behavior in soft manipulation (Puhlmann et al., 2022 ), swimming (Katzschmann et al., 2018 ), flying (Ramezani et al., 2016 ), and dynamic locomotion (Tang et al., 2020) . Each of these robots was designed manually; algorithms that design such robots, and tools for designing the algorithms that design such robots have the potential to accelerate the invention of diverse and capable robots. SoftZoo decomposes computational soft robot co-design into four elements: design representations (of morphology and control), tasks for which robots are to be optimized (mathematically, reward or objective functions), environments (including the physical models needed to simulate them), and codesign algorithms. Representationally, SoftZoo provides an unified and flexible interface of robot geometry, body stiffness, and muscle placement that can take operate on common 3D geometric primitives such as point clouds, voxel grids, and meshes. For benchmarking, SoftZoo includes a variety of dynamic tasks important in robotics, such as fast locomotion, agile turning, and path following. To study environmentally-driven robot design and motion, SoftZoo supports an extensive, naturally-inspired material set that allows it to not only simulate hyperelastic soft robots, but also emulate ground, desert, wetland, clay, ice, and snow terrains, as well as shallow and deep bodies of water. SoftZoo provides a differentiable multiphysics engine built atop the material point method (MPM) for simulating these diverse biomes. Differentiability provides a crucial ingredient for the development of co-design algorithms, which increasingly commonly exploit model-based gradients for efficient design search. This focus on differentiable multiphysical environments is in contrast to to previous work (Bhatia et al., 2021; Graule et al., 2022) which relied on simplified physical models with limited phenomena and no differentiability; this limited the types of co-design problems and algorithms to which they could be applied. The combination of differentiable multiphysics simulation with the decomposition of environmentally-driven co-design into its distinct constituent elements (representation, algorithm, environment physics, task) makes SoftZoo particularly well suited to systematically understanding the influence of design representations, physical modeling, and task objectives in the development of soft robot co-design algorithms. In summary, we contribute: • A soft robot co-design platform for locomotion in diverse environments with the support of an extensive, naturally-inspired material set, a variety of locomotion tasks, an unified interface for robot design, and a differentiable physics engine. • 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. This analysis provides insight into the efficacy of different aspects of state-of-the-art methods and steer future co-design algorithm development. 



Figure 1: An overview of SoftZoo with demonstration of various biologically-inspired designs. 2 SOFTZOO 2.1 OVERVIEW SoftZoo is a soft robot co-design platform for locomotion in diverse environments. It supports varied, naturally-inspired materials that can construct environments including ground, desert, wetland, clay, ice, shallow water, and ocean. The task set consists of fast locomotion, agile turning, and path

funding

* This work was done during an internship at the MIT-IBM Watson AI Lab. Support for this work was also provided in part by the NSF EFRI Program (Grant No. 1830901), DARPA MCS Program, MIT-IBM Watson AI Lab, and gift funding from MERL, Cisco, and Amazon. 1 Project Page: https://sites.google.com/view/softzoo-iclr-2023 

