Robust Robotic Grasping Utilising Touch Sensing

UCL COMP0029 Final Year Project Repository

This repository contains all the code and documentation mentioned in my FYP project report. The code has only been tested on a Ubuntu 20.04 dual-boot set-up. It is not guaranteed to work perfectly on other Linux distributions.

sample_pb_simulation.mp4

Prerequisites

Before compiling and testing the code from this repository, please ensure that your system has been properly configured with the following requirements:

Linux Operating System (preferably Ubuntu 20.04). The code for this project was only developed and tested on Ubuntu 20.04. It is not guaranteed to compile and run fully functionally on other operating system and devices.
Python 3.8.16
The virtualenv package for creating Python virtual environments is installed.

Deployment

Clone the repository from https://github.com/uclzzjn7/UCL-FYP.git or download a zip file containing the repository code.
```
https://github.com/uclzzjn7/UCL-FYP.git
```
If you already have Python 3.8.16 installed on your OS, please skip to step 4.

In a new terminal, run the following commands in sequence:

sudo apt update

sudo apt install software-properties-common

sudo add-apt-repository ppa:deadsnakes/ppa

sudo apt-cache policy python3.8

This should install Python 3.8.16.

Find the root path of Python 3.8.16 in your OS:
```
ls /usr/bin/python*
```
Create a Python virtual environment with Python 3.8.16.
```
virtaulenv venv --python="<path of Python 3.8.16>"
```
Assuming that the path of Python 3.8.16 is "/usr/bin/python3.8", then the command for creating a virtual environment should be:
```
virtualenv venv --python="/usr/bin/python3.8"
```
Ensure that the virtual environment is created in the root directory of the project:
```
/docs
/src
/venv
   /bin
   activate
   ...
```
Activate the virtual environment:
```
source venv/bin/activate
```

Install PyTorch:

pip3 install torch==1.13.1+cu116 torchaudio==0.13.1+cu116 torchvision==0.14.1+cu116 --extra-index-url https://download.pytorch.org/whl/cu116 --no-cache-dir

Install all required Python packages:
```
pip install -r requirements.txt
```
Run the Pybullet simulation:
```
python src/main.py
```

Using the simulation

There are several core functionalities provided by our Pybullet simulation. These functionalities are listed as either sliders or buttons on the right panel of the simulation GUI.

Manual manipulation of the robot: Adjust the six available sliders for the end effector pose (x,y,z,roll,pitch,yaw) and the slider for the gripper length (0 = gripper is completely closed, maximum = gripper is completely open).
Reset simulation: Resets all sliders and the robot (end effector position).
Get joint coordinates: Retrieves all movable joints of the robot arm and gripper.
Get object features: Retrieves the object features of the created object mesh in the simulation. Currently this function returns the 3 principal curvatures of the object, one for each dimension (width, height, depth).
Collect data (baseline): Data collection pipeline to collect data for the baseline approach. To run this function, please ensure the object_name property is set to block in the parameters.yaml file. This function will not work for any other object name since manually-selected end effector poses were only collected for block.
Collect data (proposed): Data collection pipeline to collect data for the proposed approach (MLP). To run this function, please ensure the object_name property is set to one of the following in the parameters.yaml file:
```
object_name: block1/block2/block3/bottle1/bottle2/bottle3/cylinder1/cylinder2/cylinder3
```

Training models

The baseline and proposed models for this project are trained separately from the simulation code since it requires a significantly larger amount of computing power. The relevant code can be found in the Jupyter notebooks (with the .ipynb file extension) in the models/baseline_model or models/mlp_model directories. If this code needs to be assessed, please ensure that these notebooks are run with cuda enabled, and the correct Python interpreter is selected for running the notebook (Python 3.8.16). Otherwise, there is a risk of insufficient memory allocation to training the models.

Before starting, please download the datasets folder from this link: https://drive.google.com/drive/folders/1d14Ul5YTjX-6w56OCxLbCXTLicDE3qnB?usp=sharing. This folder contains all required datasets for training the models. Extract all the contents of the datasets folder to under the src/ directory (i.e. src/datasets/).

Credits

This simulation was developed collaboratively with Jeffery Wei, a Master's 23' project student working with the Statistical Machine Learning Group at UCL.