Detection of transparent and translucent objects is of great difficulty as these objects do not possess many features and colour of their own. Initially, classical feature based techniques such as SIFT and HOG are used for object classification followed by Sliding Window over Pyramids of images and Non-Maximal Suppression for localisation. These approaches yields poor results and fails.

Deep Learning based techniques such as Faster R-CNN and SSD are then used to detect the globe in the images. These algorithms provide object detection with very high accuracy.

Problem of constructing dynamical model of complex system from first principles is of great difficulty. While fully data driven techniques exists, the models produced by these techniques does not fully explain internal dynamics. So, the dynamic model of an RC is built using a grey box technique called System Identification.

Kinematic and Dynamic Single Track Model of the RC were used as a base model and their parameters were identified using Identification experiments. For dynamic model, parameters of Tire Model were also identified.

Experiments were performed on F1/10 RC Car and VICON Motion Capture system.

This project aims at increasing the degree of autonomy in an extraterrestrial rover. This project is also is an exercise in developing a robotic system from scratch; right from sketching the rover to writing perception and motion planning algorithms. Current Mars rovers are teleoperated and heavily rely on humans in making decisions. We can reduce human reliance by incorporating the state-of-the-art mapping and navigation algorithms. This project is based on NASA’s ATHLETE (All Terrain Hex-Limbed Extraterrestrial) Rover.

Motion planning of a robot on uneven terrain is a challenging problem. This is, in part, due to the uncertain nature of the terrain; each obstacle demands different leg trajectory. A Reinforcement Learning (RL) Agent is proposed which will learn where to place the foot on the uneven terrain using Value Function Approximation. Rewards will be given based on how the episode terminates; positive rewards for successfully reaching goal location and negative reward for losing stability.

This project aims at developing fuzzy logic controller for active suspension system. A model of car with 7 Degrees of Freedom is made in Simulink and tested for various road conditions with and without controllers to stabilize the vehicle. The controllers used here are PID controller and Fuzzy Logic Controller. The results from both the controllers are compared. I have utilized MATLAB's Fuzzy Logic Toolbox to design the controller.

As the degrees of freedom of a robot increases, it becomes more and more difficult to find the solutions of inverse kinematics equations due to associated non-linearities. This project uses adaptive algorithm to solve this problem.

The idea is to do the path planning for mobile robots in a fixed environment at a very rudimentary level. Dijkstra and A-star algorithms were applied on the data set of 100+ floor plans. Exploration and final paths were visualised which gave some insight on how both the algorithms worked in different settings.

This was my first experience with a real robot. The main idea was to apply an omnidirectional static walking algorithm on a quadruped robot. The algorithm mainly focused on reducing the time between crawl and rotation gaits while maintaining the static stability of the quadruped. A gait simulator was created in MATLAB and the algorithm was first tested in it before deploying it on the actual robot (using ROS). Gait simulator was also used to simulate crawl, rotation, trot, etc. gaits by varying various parameters.