Projects

I was part of the team competing in the International Ground Vehicles Competition 2018 and 2019. We developed a fully autonomous robot which could navigate lanes while avoiding obstacles and follow global navigation goal points. The project included lane detection(on unstructured environments), localization, controls and planning modules. We were the first to qualify amongst 40+ teams and secured runners up position in the final competition. (more)

I was part of IIT Kharagpur's 8th Inter IIT Tech Meet contingent. As part of the problem statement, we designed AAROHI, an Autonomous Terrace Farming robot which can climb stairs upto 40cm in height and undertake agricultural activities like plowing, seeding, harvesting and watering/pesticide spraying. My main involvement was in Mechanical Design of the climbing mechanism, and integrating the Planning and Controls Stack for the robot. After combined efforts of 12 people for 2 months we secured GOLD in our event and and the contingent secured 2nd among 20+ other IITs in the meet. (more)

Collaborating with AGV labs IIT Kharagpur, we built a prototype of a self driving car which runs inside our campus. Challenges included urban lane detection, system integration, vehicle localization, planning and trajectory optimization, urban feature detection and many more. The car chassis was provided by Mahindra with pre-installed drive-by-wire communication setup.

Geometric and model based control algorithms were developed and tested both on a simulation platform(gazebo) and Mahindra E2O. This module was integrated with the planning module for robust trajectory generation and tracking. Methods like pure pursuit, PID, stanley, LQR and MPC were implemented. We are currently working on implementing state of the art Learning MPC based controls for increased accuracy

Road feature extraction is one of the most important tasks to be performed for prototyping a self driving car. Feature based lane feature extraction and probabilistic lane fitting methods were used on an perspective transformed frame(top view) for lane detection. Principle component analysis and lateral clustering algorithms were used for stopline detection. Unscented kalman filter(UKF) based tracking was deployed to increase robustness

Husky A200 is a rugged ground vehicle for rapid prototyping applications. Before deploying modules on Mahindra E20, All modules were tested on the A200 chasis. An onboard PC is connected to an external CPU via a static LAN network. Core modules run on the onboard PC while tasks requiring heavy processing are executed on the external CPU. GPS waypoing navigation, path tracking, path planning and vision modules have been tested on this chassis.

To aid vision based lane detection and road segmentation algorithms, a GPS tagged lane map was created of the test area. The data was stored in a custom tree like data structure for quick retrieval during a live run. Using a combination of GPS, IMU, Encoders and Camera, nearby lane data was projected back onto the camera frame and fused with a real time detection algorithm to increase robustness of lane detection and drivable region identification.