Implemented a single-shot instance segmentation algorithm from this research paper by assigning instance categories to each pixel based on location and size of instance. Achieved a mean average precision of 66.3% with a dataset consisting of vehicles, humans and animals
Implemented the YOLO algorithm for the purpose of object recognition of cars, humans and traffic lights from a dataset of more than 3000 images.
Explored the application of learning for predicting future trajectories of swarms by approximating the non-linear dynamics of swarm models. Methods from data-driven techniques like SINDy, complex learning methods like CNNs, RNNs, and NeuralODE were tested. Experiments were carried out for both steady state and transient dynamics.
Improved paths generated by probabilistic planners like RRT using shortcutting heurestics and acceleration bounding of joints for manipulator robots.
Part of the mechanical team responsible design for the chassis, gripper and other key mechanisms for the robots on SolidWorks and performed kinematic analyses of the designed mechanisms. In the competition, we won the team award for the 'Best Use of Matlab'. We had used Matlab to establish the inverse kinematic equations for the mobile robots which were using Mecanum and Omni wheel configurations.
Designed a novel Continuously Variable Transmission (CVT) incorporating positive drive. The design enables for good fuel economy for high torque applications like in trucks and buses.