Currently working on ADAS (Advance Driver Assistive System) which features FCW(Foreward Collision Warning), LDW(Lane Departure warning), BSD(Blind Spot Detection) , etc.
Specialized in development of Computer Vision algorihtms for object detection, semantic segmentation, instance segmentation ,Tracking, sentimant anyalysis, face recognintion using deep learning and
machine learning learning techniques. Well Versed in deep leaning framework like, caffe, tensorrt, pytorch , tensorflow and keras.
Technicalized in development of algorithms on vision, motion planning, calibration for robotic automation in warehouse and construction industry applications. Proficient in programming using libraries like OpenCV, PCL using C++ and Python language. Have extensive experience working on industrial robotic manipulators like Barrett WAM, Univeral Robots (UR5, UR10).An expert at design and implementation of robotic tasks in real world and simulation using MoveIt, Gazebo platforms using ROS architecure.
ADAS System development.
Design and implementation of algorithms for robotics and computer vision in automation of warehouses. Developing deep neural network algorithms for object detection and segmentation using Caffe and Tensorflow libraries. Creation of URDF, SDF, XACRO files for simulation of robots. Creating client presentation and demos on vision and robotics. Training and imparting knowledge to new recruits for working with robots.
This Projects Implements the semantic segmentation using deep learning algorithm which gives the pixel wise probabilities for the 40 objects and background.ResNet architecture was modified which accelerate the training and increase the accuracy.Algorithm was developed on both Tensorflow and Caffe deep learning libraries.
View ProjectThis project implements 3D model matching using SUPER4PCS, 3D models are artificially generated using a combination of primitive shapes such as cuboid, cylinder, sphere. Principal component analysis is applied on the 3D model of the object to compute the appropriate normal for the object against which the robot has to approach the object. The touch point for suction is computed using the scene 3D points for the target object.
View ProjectThis project implements the motion planning for robotic manipulators mounted with two finger gripper or suction based grasping system. Implementation of motion planning using MoveIt as planning library is done with stage wise improvements. At first in Amazon picking challenge 2016, motion plan for default orientations based on the object's location wrt the bins walls is implemented. Motion planning is improved to pick objects using a gripper with objects placed at any orientation in a rack. Further improvisation is achieved to pick an item placed with any orientation at any corner of a bin using suction. Collision avoidance was implemented using collision plates for storage rack walls and Octomaps for collision avoidance with nearby non-target items.
View Video 1 View Video 2 View PaperThis project deals with the calibration (estimation of transformation) between the 3D sensor and fixed robot reference frame. 3D points are collected wrt to the 3D sensor frame and wrt fixed robot reference frame. The transformation between the two sets of points is computed using SVD.
To estimate the object's location wrt the walls of bins in a storage rack it is necessary to know the corners of the every bin in the rack. The algorithm recognizes the grid pattern in the rack and computes the bin corners for all the bins.
This project implements an algorithm for computing the grasp position on novel objects without any prior information of the objects. The algorithm collects the 3D scan of an object, randomly selects the seed points from the 3D scan. For each seed point a nearest neighbour search is performed to collect all the points that are within a sphere having diameter equal to the opening width of the gripper of the robot. Surface fitting is performed using the points around the seed points, the surfaces having radius of curvature less than half the width of gripper is chosen as a candidate grasp position. After surface fitting and selection of candidate grasps for all the seed points, the candidate grasps are filtered for obtaining the grasping position that is closer to the central position of the target object
View VideoMember of team IITK-TCS which participated in Amazon Robotics Challenge, held in RoboCup 2017, Nagoya, Japan. Won 3rd place in pick task, 5th place in stow task and 4th place in final round out of 16 teams in the competition.
IITK-TCS Website Github Repository View ResultsMember of team IITK-TCS which participated in Amazon Picking Challenge, held in RoboCup 2016, Leipzig, Germany. Achieved 5th place in stow task, 10th place in pick task out of 16 teams in the competition.
IITK-TCS Website View Pick Results View Stow ResultsAwarded gold medal for outstanding academic performance in convocation 2016 at IIT Jodhpur.