This page contains detail descriptions of projects I did in past with results and report on the project and links to github for respective projects.

List of Projects:

  1. Transfer Learning in Reinforcement learning with application to Robotics.
  2. Robot Path Planning with differential constraints.
  3. Pedestrian Tracking from Mobile Robot.
  4. Predicting migration location of White Stock using LSTM.
  5. Driving Aquatic agent autonomously on Pond

Abstract and links to each project (click the project title or image) is given below.

1. Transfer learning In Reinforcement learning with application to Robotics. (Report, Git-repo)

Under Prof. Doina Precupp.

Autonomous robots have achieved high levels of performance and reliability at specific tasks. However it is important for a robot agent to be able to adapt to the new environment and learn varying tasks. In Reinforcement learning agent learns by interacting with the environment and gathering data. Therefore learning different tasks in isolation can be very expensive for a physical agent both in terms of computation and actual physical cost of the Robot. Transfer learning can be used in such cases to learn in simulated environment and using the learned knowledge on actual physical robot to avoid damage as well as to speed up vanilla RL algorithms. This report focuses on understanding transfer learning problem in reinforcement learning domain and applying it to robot navigation task.


2. Robot Path Planning with Differential Constraints. (Demo, Report, Git-repo)

Under Prof. David Meger.

Even the simplest robot model is subject to differential constraint. In robotics most problems involve differential constraints that arise from kinematics and dynamics of the robot. In order to plan a collision free path for robot which it can successfully follow its important to consider these constraints while planning. This project studies a very popular randomised path planning technique called RRT-Planning and apply it for a simple robot model with differential constraints. In order to deal with differential constraints a new sampling technique is studied in the report which is based on feasibility sets. Report also discusses in details theory of reachability sets and discusses reachability sets for simple car motion. Properties of RRTs and how the differential constraint problem affect those properties are also discussed in the report. Finally report also discusses the scope of improvement in the studied algorithms. (Please click on image for the video, link in heading for report.)


3. Pedestrian Tracking from a Mobile Robot. (Report, Git-repo)

Under Prof. Kaleem Siddiqi

The ability to track moving objects in a video made with a moving camera is a task that humans perform with remarkable ease and triviality (and they actually do so with two moving cameras, the eyes), yet as a problem for a computer vision system it turns out not to be so easy. This problem is characterised by a dual character of two independent motions. On the one hand, we have the movement of the moving camera (e.g., on a the robot) in the environment. On the other hand, we have the external motion of objects in the environment (e.g., pedestrians). This project deals with tracking pedestrian using a camera mounted on a mobile robot.



4. Predicting Migration Location of White Stork using LSTM (Report, Git-repo)

Under Prof. Joelle Pineau

White storks are migratory birds that have a diverse migratory pattern that often causes varied wintering locations among juvenile birds within the species. Birds that winter in areas with a higher human population tend to behave differently from those that don’t. Since they are affected by their ecological system, it would be beneficial to map out their widespread migration patterns. A better understanding of what areas are affecting the birds in their migration routes would help in knowing where to concentrate conservation efforts. We develop several algorithms to help us to help us predict migratory patterns of these birds. First we cluster the birds in order to find out which birds are travelling together so we can map out maps for these clusters. Since we are working with time series data we then use LSTM and compare it with a regression model to predict future paths.

5. Driving Aquatic agent autonomously on Pond (Demo, Git-repo)

Under Prof. Micheal Jenkin (York University)

During my summer internship at York University I programmed an aquatic agent to drive autonomously on the pond. Project involved localising this aquatic agent on a water body using particle filter where correction feedback on position of the robot was obtained using off-board camera and on board IMU. Position of the robot was tracked using Homography and coloured object tracking. Image shows the user interface I programmed to control the agent. Click on the link in the heading for the video.