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

This repository contains code for the paper Soft Robots Learn to Crawl: Jointly Optimizing Design and Control with Sim-to-Real Transfer.

It provides the code for model order reduction, co-optimiztion of design and control, and testing the optimized design-control pairs in simulation.

Installation

Our code relies on Docker and the docker wrapper cpk to manage its dependencies.

To use this codebase, follow these installation steps:

1. Intstall Docker.
2. Install cpk: python -m pip install cpk
3. Clone this repository.
4. From the top level directory, run cpk build to build the docker container.

Additionally, this codebase uses Weights and Biases for logging and visualization, which is free to use for academics. Running a training job will prompt you to login, create an account, or not visualize the results. If you wish to avoid manual logins, place your wandb .netrc file in the assets/wandb_info directory.

If you only want to visiualize the pretrained model, then creating a wandb account is not required.

Testing the Learned and Baseline Models

We provide code to visualize a pretrained model and our baseline.

• To visualize the pretrained model, run: cpk run -M -f -X -L viz_pretrained
• To visualize the baseline, run: cpk run -M -f -X -L viz_baseline

Running a Co-optimization Experiment

The training code is run in two pieces: a training script and an evaluation script which runs concurrently. Training is CPU intensive. We ran this experiment with 96 paralell SOFA environments for training on a 32-core AMD EPYC 7502. You can adjust the cpu load by editing the config file configs/coopt.yaml.

• To launch a training job, run: cpk run -M -f -n train -L train
• To launch the eval job, run: cpk run -M -f -n eval -L eval

The eval job will search for checkpoints in the exps directory and generate videos and evaluation performance. All logs and videos will be visible on wandb in a newly created evolving-soft-robots project.

Performing Model Order Reduction

This repository contains the reduced order models used in this paper, as well as code to create your own reductions.

• To launch a reduction, run: python scripts/launch_reduction_block.py

This will launch a grid search over the two tolerances for the reduction. The reduction takes several hours to complete and requires large amount of RAM (~200GB).