As part of my MSc in Robotic Systems Engineering at RWTH Aachen University, I participated in a challenge for the Reinforcement Learning and Learning-Based Control (RLLBC) course. Using the existing Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm, I taught an autonomous car to navigate a race track and avoid safety cones. My implementation achieved the second-highest score in the class, surpassed only by the course instructor's agent. Note: The code for this project is not available as it was part of a class. (Project Link)

In this project, my main goal was to use Reinforcement Learning to learn a policy for the task of position tracking in quadrotor. I used Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm with a 3 layered Neural Network for both Policy and Critic Networks. I used Position Difference, Translational Velocity, Angular Velocity, Rotation Matrix for states and actions were RPMs (motor speed) for each motor. The resulting policy seems to track the target position with an error of 0.06 units. (Project Link)

Through this project my aim was to test various Reinforcement Learning Algorithm for continouos control problem before they could be applied in research work. I used OpenAI Gym's Bipedal Walker Environment in this project. Due to the simplicity of this environment, the implementation and debugging of RL algorithm can easily be done. I used this environment for testing Neuroevolution of augmenting topologies (NEAT), Proximal Policy Optimization (PPO), Deep Deterministic Policy Gradient (DDPG), DDPG with parameter noise, Twin Delayed Deep Deterministic Policy Gradient (TD3).

Neuroevolution of augmenting topologies (NEAT) (Project Link)
Proximal Policy Optimization (PPO) (Project Link)
Deep Deterministic Policy Gradient (DDPG) (Project Link)
DDPG with parameter noise (Project Link)
Twin Delayed Deep Deterministic Policy Gradient (TD3) (Project Link)

This project tries to solve the control problem of Cartpole Balancing using Reinforcement Learning Methods. The environment used in this project is from OpenAI Gym's Cartpole Environment. It uses discreet action space (go left or go right) to control the cart which inturn balances the pole. I have used PPO and DDPG for this project

Proximal Policy Optimization (PPO) (Project Link)
Deep Deterministic Policy Gradient (DDPG) (Project Link)

This project is based on the paper named "Deep Reinforcement Learning with Double Q-learning" (DDQN). It uses a Convolutional Neural Network to learn a Q function. This Q function can then be used to choose action which gave have best Q Value. For exploration, we sometime choose a random action, thus forming a Epsilon Greedy Policy. (Project Link)

This project was a part of my final year thesis. In this, a Convolutional Neural Network (EfficientNet) is used to learn semantic information from an RGB image of an Urban Scene. I used a method called tranfer learning to reduce computational burden. (Project Link)

This project was a part of my final year thesis. In this, a Convolutional Neural Network (EfficientNet) is used to learn depth information from an RGB image of an Urban Scene. During the training process, stereo image pair are used. During the inference, only one image is used to predict the Depth/Disparity map from an image. Tranfer learning is also used to reduce computational burden. (Project Link)

This project was a part of my Internship Project. It is uses Haar Cascade and Convolutional Neural Network to determine if a Person has the Covid Mask On or not. (Project Link)

In this project we use OpenCV and HaarCascade for face detection.(Project Link)