Cloud Folding

Code for the paper "Learning to Fold Real Garments with One Arm: A Case Study in Cloud-Based Robotics Research" (website). This repository is a fork of Google PyReach. Changes from the parent repository are summarized below.

Fork Details

Almost all code changes are in the pyreach/tools and pyreach/examples sub-directories.

In pyreach/examples, we add several agents for the benchmark_folding_v2 PyReach Gym environment (filenames begin with bair_), where each agent executes a different algorithm from the paper (e.g. CRL or KP). These can be run with python -m pyreach.examples.[filename] as described in the PyReach README below.

In pyreach/tools we implement code to support the execution of the agent code above. For instance,

• pyreach/tools/basic_teleop.py provides a GUI for human pick-and-place teleoperation.
• corner_pulling.py implements key parts of the KP algorithm.
• the pyreach/tools/inv_dynamics/ directory has data processing, model training, and model inference code for the IDYN and CRL algorithms.
• the pyreach/tools/reach_keypoints/ directory has model training and inference code for training LPAP and KP. This directory is a copy of hulk-keypoints on Github, which should be directly cloned from the source to avoid issues if you wish to re-train models for those algorithms.

To run our algorithms on Reach successfully, the following must be true:

• You will need credentials to access Reach, which is currently invite-only.
• You will need to merge upstream changes from PyReach, since PyReach is constantly updated and this repository is only guaranteed to be compatible with the current version of PyReach. If this leads to nontrivial conflicts, feel free to raise a Github issue and we will try to resolve it.
• You can get our trained models, collected datasets, etc. from the project website: copy them to the data/ directory or an otherwise appropriate location for the agents to read them correctly.

The original Google PyReach repository's README follows.

PyReach README

Disclaimer: this is not an officially supported Google product.

PyReach is the Python client SDK to remotely control robots. Operator credentials will be approved by invitation only at this time

Supported Platform

Ubuntu 18.04, Ubuntu 20.04, and gLinux.

Build

git clone https://github.com/google-research/pyreach.git
cd pyreach
./build.sh

Getting Started

If build.sh runs successfully, PyReach is ready to be used. It is supported, but not required to install PyReach into the system path.

Step 1. (Only the first time) Login to the system.

cd pyreach
./reach login

# Follow the login instructions.

Step 2. Connect to a robot.

In a new shell session.

cd pyreach
./reach ls
./reach connect <robot id>

Step 3. View camera image

In a new shell session.

cd pyreach
./reach-viewer

Step 4. Run pendant tool.

In a new shell session.

cd pyreach
./reach-pendant

Step 5. Run example agent.

In a new shell session.

cd pyreach
source setenv.sh
python -m pyreach.examples.pyreach_gym_example

Logs

By default, all the client logs are saved under:

$HOME/reach_workspace

Install

Optionally, PyReach can be installed as a pip package:

cd pyreach
pip install .

Once PyReach is installed, the command "reach", "reach-viewer", and "reach-pendant" can be access directly through command line.

Uninstall

To remove the PyReach pip package:

pip uninstall pyreach