VISHAC - Efficient Visuo-Haptic Object Shape Completion for Robot Manipulation

This repository includes code for Efficient Visuo-Haptic Object Shape Completion for Robot Manipulation paper.

Authors: Lukas Rustler, Jiri Matas, and Matej Hoffmann Maintainer: Lukas Rustler, [email protected]

Contents

• Licensing
• Installation
• Data
• Run
• Real Setup
• Publication / citing
• Acknowledgements

Licensing

The code created in this work is under GNU-3 licence. For licensing information about otherused repositories see: LICENSE, and the original repositories:

• ODIO-URDF
• MuJoCo-ROS
• Kinova_Mujoco
• V-HACD
• IGR
• ros_kortex
• YOLOv7
• orocos

MuJoCo is a trademark of DeepMind (https://mujoco.org)

Installation

The code is tested on Ubuntu 20.04 with ROS Noetic. It will run only on unix-based systems, i.e., it will not work in windows.

There are two options to install/use the code: 1) Docker (recommended) and 2) manual installation.

Docker

The image is available at: Docker hub

Steps to use:

1. Install Docker from https://docs.docker.com/engine/install/

2. Install nvidia-docker from https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html#docker

   • for GPU

3. Create a new folder (e.g., VISHAC), clone this repository, and rename the downloaded folder to VISHAC

   cd some_path/VISHAC
   git clone https://github.com/ctu-vras/vishac.git
   mv vishac src
   

4. Use the provided script run.sh with xhost local:root

   • the script takes 2 mandatory parameters:
     • name of the container
       • string
       • recommended: vishac
     • path to link inside the container
       • string
       • should be: some_path/VISHAC from the previous step
   • optional parameter 'cpu' can be use to run the container in CPU only mode
   • e.g., xhost local:root && ./run.sh vishac /home/my_user/VISHAC
     • If you run it non-linux enviroment, running over SSH or it doesnt work, run: vnc && start-vnc-server.sh inside the docker container. This will open VNC server inside the container, and you can access it through your browser at: http://ip_address:6080. Where ip_address is localhost if you run it on local machine, or ip address of a remote machine.

5. Now, when you are in the docker environment, build the workspace:

   cd some_path/VISHAC
   source /opt/ros/noetic/setup.bash
   catkin init
   catkin build
   source devel/setup.bash
   

Manual installation

Manual installation is also possible. Ubuntu 20.04 is recommended. Install ROS Noetic, Python 3 with all dependencies (mainly cv2, trimesh, open3d, numpy, pytorch; see Dockerfile for more information). The build instructions are the same as for Docker.

Data

Data are available at OSF.io platform.

• To run new experiments or repeat, you only need to download GT_meshes and extract them to kinova_mujoco/GT_meshes.

• If you want to use data collected during our experiments, meshes, point clouds, logs, rosbags, metrics and numpy files are available to download as well. Also, experiments files used are available in exps directory.

When running the pipeline, all data will be stored at: some_path/VISHAC/src/shape_completion/data. Completed meshes, collected point clouds, corresponding numpy files and rosbags are saved after every touch. In addition, after using evaluation scripts (see below) Python Pickle files will be created.

Run

The easiest way how to run experiments is to:

1. prepare experiment file in exps directory. See example.json or single.json files.
   • The json file should include:
     • objects: [["object name 1", ..., "object name n"], [...]]
       • object names are taken from names of the objects in GT_meshes. Without file extension.
       • each element can contain up to 5 objects
       • each sub list is one experiment
     • origins: [[origin 1, ..., origin n], [...]]
       • origins are in the form of list [x, y, z]
       • one origin for each object
     • repetitions: [int, ...]
       • number of repetitions for given experiment
     • reconstructions: [int, ...]
       • number of touches in given experiment
     • max_time: [float, ...]
       • max time for each experiment
2. run rosrun shape_completion do_experiments.py -s some_path/VISHAC/src/shape_completion/data/exps/your_experiment.log
   • this will run the experiment defined in your exp file
   • the results will be written in log directory
   • in the end, evaluation script is run to compute Jaccard similarity and Chamfer distance

If you want to run everything by yourself:

• Run the simulation run_simulation object_name_1,...,object_name_n [[origin 1,...,origin_n]] 'false' 'true' 'true' FPS free_space number_of_points 0
  • where:
    • FPS - whether to use Farthest Point Sampling, true or false
    • free_space - whether to use free space, true or false
    • number_of_points - how many point to use in each IGR iteration, we used 500
  • run_simulation is alias for run_simulation.sh.
    • if you use docker, the alias should already exist
• run the main script rosrun shape_completion main.py -r number_of_reconstruction -t max_time
  • where:
    • number_of_reconstruction - number of touches + 1
    • max_time - desired maximum running time of the pipeline
• If you want to evaluate the results, run rosrun shape_completion evaluation_utils.py name_of_log_file
  • where name_of_log_file is name of log created with main.py saved in logs directory.

Real Setup

This repository does not include drivers for the real-world devices. However, the code is unchanged and can be used with real robot. The code contains parameter real_setup, which control behaviour of some functions.

You can get inspired from real.launch, that was used for real experiments with Kinova Gen3, with Robotiq 2F-85 gripper and Intel D435 camera.

Publication

This repository is connected to the following publication:

Rustler, L., Matas, J., & Hoffmann, M. (2023). 'Efficient Visuo-Haptic Object Shape Completion for Robot Manipulation '. In Intelligent Robots and Systems (IROS), IEEE/RSJ International Conference on. [DOI]

If you use the code or data in this repository, please cite this article.

Video

A video accompanying the article is availabe on [youtube].

Acknowledgements

This work was supported by the OP VVV MEYS funded project CZ.02.1.01/0.0/0.0/16 019/0000765 “Research Center for Informatics”. L.R. was additionally supported by the Czech Technical University in Prague, grant no. SGS22/111/OHK3/2T/13.

We also thank to authors of Act-VH. Last but not least, we thank Robotology community for help the with Docker image.