HenryWJL/universal_manipulation

An upgraded version of universal manipulation interface (UMI)

Universal Manipulation

This is a variant of universal manipulation interface (UMI). Concretely, we made the following changes:

• Other than GoPro cameras, users can now utilize their mobile phones as visual sensors and IMU processors.
• Details of camera calibration and camera-imu calibration are included.

Hardware Requirements

UMI Gripper

Hardware Guide

Mobile Phone with PilotGuru

Download PilotGuru on your mobile phone.

Installation

1. Install system-level dependencies:

sudo apt install -y libosmesa6-dev libgl1-mesa-glx libglfw3 patchelf

2. Download this repository:

git clone https://github.com/HenryWJL/universal_manipulation.git

3. Create a conda environment with all the dependencies.

conda env create -f conda_environment.yaml

Usage

1. Calibration

Follow calibration instructions.

2. Data Collection

(1) Record mapping video

Print out the ArUco marker in aruco_mapping.pdf and place it in the real-world working space (e.g., if you want to manipulate objects on a table, place it on that table). Then record a video (1 min) using PilotGuru.

(2) Record demonstrations

Mount your mobile phone onto the UMI gripper and hold the gripper to manipulate objects with PilotGuru turned on. Make sure your motions are not too fast in order to avoid motion blur.

3. Dataset Generation

Now your mobile phone should contain a directory dubbed PilotGuru that stores several videos and IMU data recorded by PilotGuru. Move it under universal_manipulation directory and run:

python run_slam_pipeline.py -l PilotGuru