SUSHANT KUMAR

STAIR CLIMBER

Summary

This is a project that is aimed at making a bot find, align, and then climb stairs. The environment was will be built using ROS and Gazebo. Then an RL agent will be trained using the simulation data, which will include RGBD images. Then the motors of the bot will be actuated according to the predicted actions. Once the model is trained and works well on simulation, we will do a sim2real transfer. Then it will be implemented hardware and we will test in the real world scenario. The data used for training is collected by the members of IvLabs. The Labelling and preprocessing is done by the team. Yet we are reading the diffrent SOTA algorithms and trying them out on a single environment.

OSWALT

Abstract

OSWALT - Omnidirectional Spherical Wall Traversing Robot. Inspection of ducts and chimnies in industries has always been a problem open for researchers. Here we provide a solution for inspection of ducts and places where reaching physically is not just difficult but at times impossible. This bot can climb walls at any angle and get pictures. It has a unique gimble mechanism that keeps it stable while it climbs changes direction or performs complex maneuvers.

Cart Pole Solver

Solution Overview

The Cart Pole environment is a fundamental problem in control engineering. As it is an easy to train environment with fewer state parameters, I have tested all the discrete RL based algorithms on this environment. There is another similar environment of Inverted Pendulum which has continuous action space and can be trained easily. I have used these environments for implementation of various algorithms and further I am planning to train the other environments like Lunar Lander, Biped, etc.

Single Digit Classifier

Summary

A deep neural network was designed and trained using just the concepts of NumPy without the use of any Framework to classify handwritten single digits from 0-9. Then a CNN was trained using PyTorch, to improve the accuracy. The virtual drawing pad that was built using thresholding techniques in Open CV was then used to give real-time input from the webcam, and the model was able to recognize the number drawn.

Model architecture

3DOF ARM SIMULATION

Summary

This was a mini-project to understand the forward kinematics of a 3 DOF robotic arm. It was an implementation project for the hands-on experience of the Introduction to the robotics course of Stanford. It is designed to move in 3-D space. This project is my gateway into the field of kinematics. The rotation and translation matrices were written from scratch and implemented. The user gives in the input angles of the arm wrt the previous link and the arm moves into the specified configuration moving one link at a time.

DRAW IN AIR

Summary

This project was aimed at understanding the basic concepts of OpenCV. In this project I read about the RGB and HSV colour spaces. It helped me get hands-on experience and understanding the advantage of HSV colour space over RGB. Also the concepts of countours and centeroid was understood and implemented. I learnt colour thresholding and using of trackbars. Then mask was created to get the two diffrent images one with the black background and white colour drawing and the one with the webcam feed. This can be used for Interactive teaching without the use of expensive drawing gadgets.