# Soft Robots Learn to Crawl: Jointly Optimizing Design and Control with Sim-to-Real Transfer

**Authors:** Charles Schaff, Audrey Sedal, Matthew R. Walter  
**arXiv:** [2202.04575](https://arxiv.org/abs/2202.04575)  
**Date:** 2022-02-09  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2202.04575v1/

## Key points

- Comprehensive framework for design and control of soft legged robots
- Focus on simulation, co-optimization, and sim-to-real transfer
- Soft robots possess compliance, enabling challenging behaviors
- Interplay between mechanical design and control system is crucial
- Co-optimization addresses coupling between design and control
- Challenges in creating accurate yet fast simulation environments for soft robots
- Combination of finite element simulation and model order reduction techniques provides efficiency and accuracy
- Utilizes reinforcement learning based approaches for co-optimization
- Successful optimization, construction, and sim-to-real transfer of crawling soft robots demonstrated
- Learned robot outperforms expert designed crawling robot
- Contributes to advancing field of soft robotics by providing robust framework

## Summary

This work presents a comprehensive framework for the design and control of soft legged robots, focusing on their simulation, co-optimization, and sim-to-real transfer. Soft robots possess a unique characteristic known as compliance which enables them to exhibit behaviors that are challenging to program using traditional methods. To fully exploit this mechanical intelligence, it is crucial to consider the interplay between the robot's mechanical design and its control system. Co-optimization offers a promising approach for generating sophisticated soft robots by effectively addressing this coupling. The complex dynamics of soft robots pose challenges in creating simulation environments that are both accurate enough for sim-to-real transfer and fast enough for contemporary co-optimization algorithms. In this study, the authors demonstrate that combining finite element simulation with recent model order reduction techniques can provide the necessary efficiency and accuracy to successfully learn effective design-control pairs for soft robots that can be transferred to real-world applications. The proposed framework utilizes reinforcement learning based approaches for co-optimization and showcases successful optimization, construction, and zero shot sim-to-real transfer of several crawling soft robots. Remarkably, the learned robot outperforms an expert designed crawling robot highlighting the ability of this approach to generate novel designs that excel even in well understood domains. Overall, this work contributes significantly to advancing the field of soft robotics by providing a robust framework for designing and controlling these unique systems. The combination of simulation, co optimization and sim to real transfer opens up new possibilities for developing highly capable soft legged robots with improved performance and adaptability.