Pose Estimation for a Robot Manipulator

---

Project description

This project proposes a solution for position estimation of a single industrial manipulator using learning methods.

The model uses Region Based Convolutional Neural Network (R-CNN) to track robot keypoints which allows to predict the pose of robotic manipulator. The keypoints are defined as:

• A robot joint;
• The joint's "head";
• The joint's "tail".

The model prediction result examples:

Dataset

Dataset was created and preprocessed using Roboflow. Available in a Keypoints.v2i.yolov5pytorch.zip archive.

Example of labeled image from dataset

Model

The model is created, saved, and tested in the Pose_estimation.ipynb.

Jupiter Notebook Demonstration.ipynb is used for demonstration purposes, where both the dataset and model are loaded in to illustrate the predictions.

Installation guide

Prerequisites

• python v3 is required (v3.7+). Its installation guide can be found here
• PyTorch Vision library for object detection The necessery files can be obtained from the repository directly.
• Additionally, the following python libraries are required:

torch
torchvision
albumentations
cython
pycocotools
matplotlib
numpy

These can be installed using pip

Further installation and launch

1. Clone the repository:

git clone https://github.com/mirnanoukari/Pose-Estimation.git

2. Run Pose_estimation.ipynb to create and save the model;
3. Run Demonstration.ipynb.