hardware_development.md

Developing on the Hello Robot Stretch

We use a multi-system setup for controlling the Hello Stretch, where low-level control runs on the "Robot" and large neural nets are evaluated on a "Workstation" -- which here refers to a local computer with a GPU. For best performance, your Workstation should be on the same wireless network as the Robot, preferrably with an ethernet connection to the router for lowest latency.

Installation

• Setup instructions on Robot
• Setup instructions on Workstation

Usage

On Robot: Launching the Hello Stretch hardware drivers

roslaunch home_robot_hw startup_stretch_hector_slam.launch

On Workstation: Interacting with the robot via the client

See examples for common API usage.

You can also start an interactive command line interface (CLI) with:

python -m home_robot_hw.remote.interactive_cli

Appendix

Syncing code between Robot and Workstation

Let ROBOT_IP store the robot's IP and let WORKSTATION_IP store the workstation's IP. If your local network doesn't have access to internet we recommend using rsync with --update flag to sync your code changes across the machines. Usage:

rsync -rvu /abs/path/to/local/home-robot $ROBOT_USER@$ROBOT_IP:/abs/path/to/remote/home-robot

The above command will do a recursive updating of changed files while echoing a verbose output.

Collecting data on the real robot

We provide scripts to collect data in H5 format using Recorder class. Follow the instructions for recording these files. If your application needs more/fewer data-sources, sub-class Recorder and over-ride the save_frame method Collect the data through teleoperation.

python collect_h5.py --task_name TASK_NAME  --dir_path DIR_PATH

This will save the teleoperation files at DIR_PATH/TASK_NAME-{iteration}/{datetime}.h5.

Turn on the controller and make sure it is connected to the robot (top two blue lights should be on). To give trajectory commands through the Xbox controller, open a separate terminal and:

rosrun joy joy_node

Refer to official hello robot keybindings to learn how to operate Stretch with the controller. We also provide a Dataloaders to load data into torch (WIP).