Skip to content

AbhishekRS4/cognitive_robotics_grasping_manipulation

 
 

Repository files navigation

Coupling between Perception and Manipulation: Learning to Grasp Objects in Highly Cluttered Environments

Three scenarios: (left) isolated scenario, (center) packed scenario, and (right) pile scenario.

Assignment Overview

Service robots typically use a perception system to perceive the world. The perception system provides valuable information that the robot has to consider for interacting with users and environments. A robot needs to know how to grasp and manipulate objects in different situations to assist humans in various daily tasks. For instance, consider a robotic task such as clear table. Such tasks consist of two phases: the first one is the perception of the object, and the second phase is dedicated to the planning and execution of the manipulation task. In this assignment, you mainly focus on the topic of deep visual object grasping and manipulation.

The main goal of this assignment is to make a coupling between perception and manipulation using eye-to-hand camera coordination. Towards this goal, we have developed a simulation environment in PyBullet, where a Universal Robot (UR5e) with a two-fingered Robotiq 2F-140 gripper perceives the environment through an RGB-D camera. The experimental setup for this assignment is shown in the following figure. This setup is very useful to extensively evaluate different object grasping approaches.

Our experimental setup consists of a table, a basket, a UR5e robotic arm, and objects from YCB dataset. The green rectangle shows the robot's workspace, and the camera indicates the pose of the camera in the environment. Synthesis RGB and depth images, together with a segmentation mask are shown on the left side of the figure.

We are pursuing three main goals: (i) learning about at least two deep visual grasping approaches, (ii) evaluating and comparing their performances in three scenarios: isolated, packed, and pile (see a video of each scenario above); (iii) investigating the usefulness of formulating object grasping as an object-agnostic problem for general purpose tasks. You can also use this setup to develop your final course project.

In this assignment, we capture an RGB-D image of the scene and pass the image to a deep convolutional neural network to obtain pixel-wise grasp configuration in terms of grasp quality, grasp angle, and grasp width. To make it clear, we visualize the output of the GR-ConvNet network for a given image:

Outputs of the GR-ConvNet network for a given image

The best grasp configuration is then selected and then, we convert the grasp pose from pixel space to the robot's workspace (x, y, z, roll, pitch, yaw). We finally instruct the robot to perform a clear table task by grasping and manipulating the target object from the table to the basket. A particular grasp is recorded as a success if the object is inside the basket at the end of the experiment. An experiment is continued until either all objects get removed from the workspace, or four failures occurred consecutively. Note that, the system automatically reports a summary of the obtained results in the “results” folder, and the prediction of network is visualized and saved in the “network_output” folder.

Requirements

Ensure you are running Python>=3.6.5 and import the required libraries by running:

cd ~
git clone https://github.com/SeyedHamidreza/cognitive_robotics_manipulation.git
cd ~/cognitive_robotics_manipulation
python3 -m pip install --upgrade pip
pip3 install -r requirements.txt

It will install a set of packages, including: numpy, opencv-python, matplotlib, scikit-image, imageio, torch, torchvision, torchsummary, tensorboardX, pyrealsense2, Pillow, pandas, matplotlib, pybullet

cd ~
gedit .bashrc

and then add the following lines at the end of your .bashrc file

#This line is necessary for MoveIt! and Pybullet, otherwise the robot seems broken
export LC_NUMERIC="en_US.UTF-8"

close all your terminal and open one.

How to run experiments

We can perform a simulation experiment by running the 'simulation.py' script. As shown in the following image, we can perform experiments in three different grasping scenarios, including isolated, packed, and pile scenarios:

cd ~/cognitive_robotics_manipulation
python3 simulation.py --scenario=pile --network=GR_ConvNet --runs=10 --save-network-output=True
  • Run 'simulation.py --help' to see a full list of options.

    • --runs=10 forces the system to run 10 experiments

    • In the environment/env.py file, we have provided a parameter namely SIMULATION_STEP_DELAY to control the speed of the simulator, this parameter should be tuned based on your hardware.

    • After performing each experiment, a summary of the results will be visualized and saved in the results folder.

    • Furthermore, you can check the output of the network by setting the --save-network-output=True. The output will be saved into the network_output folder

Integrating a new model

You need to add your trained model into the "trained_models" folder. You can check the code (simulation.py and grasp_generator.py) to see how we integrate and use the GR-ConvNet model.

References

CONTACT INFORMATION

  1. Please use the following email addresses if you have questions or want to contribute to this project:
  2. check out IRL-Lab website for more information about other projects.

TODO

  • Add a param to save/not save the results
  • Objects get disappear after several runs, check the potential reasons.
  • IK should be improved -- especially for the lower-right position, it does not work properly

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 99.7%
  • Shell 0.3%