Since August,2011 I am working as a Research Assistant at Robotics Research Center, under the guidance of Dr. Madhava Krishna.

My primary research areas are:

Coordinated UGV-MAV Exploration for 2D Augmented Mapping

Arpose Arpose

In this project we proposed exploration technique involving multiple pairs of UGVs and MAVs. Each UGV is known to be equipped with a laser range finder and a monocular camera, whereas MAVs carry a monocular camera and a AR Tag. Both UGV and MAV are running a SLAM algorithm. UGV creates a 2D occupancy grid map using laser whereas MAV creates a sparse pointcloud map using a monocular camera.

Due to erratic odometry reading form MAV, a visibility constraint was introduced. Using its camera UGV tracks the AR tag attached to drone. Due to accurate state estimation of MAV, the pointcloud map created was UGV was made on correct scale. Maps received from UGV and MAV is then used for creating Augmented 2D maps.

Strategy for exploration is formulated as 0-1 Integer Programming(IP)problem. IP provides the scope of adding constraints seamlessly and makes the formulation elegant. The details about objective function and various constraints used for formulation problem is mentioned in detail in the paper.

The whole framework was designed using ROS and OpenSceneGraph. The framework was tested on multiple simulation environment and on limited hardware(1 Pioneer P3DX, 1 turtlebot and a Parrot Ar Drone).

I presented initial results for this work at poster session in AAMAS,2013 and I'll be giving an oral presentation for this at IROS,2013. Paper can be accessed from publications section.

Besides me two other students who contributed in this project are Aravindh Mahendran and Nikhil Soni.

A video demonstrating simulation can be seen in this video.

Distributed and Asynchronous Multi Robot Exploration Using Max-Sum

In continuation to previous work, this project involves multi robot exploration for a distributed and asynchronous system. A multi robot system is essentially asynchronous since different robots can realize there allocated frontier/goal position at different time instants and can continue exploration without waiting for other agents to reach there goal positions.

Besides asynchronicity,distribution is another important characteristic property of the system. In distributed setting each robot/agent is responsible for its own allotment, making system more robust than a centralized system since a distributed system does not suffer from problem of central point of failure and can handle breaking of communication robustly.

The problem of task allocation was formulated as Distributed Constraint Optimization problem (DCOP). The constraints among agents is represented using Factor Graphs and problem is solved using Max-Sum Algorithm.

Max-Sum is an iterative message passing algorithm where agents exchange information among themselves and converge to a solution which is optimal for whole team. Formulating the problem in this structure assured the required distribution and asynchronicity plus it also provided the scope of addition/removal of new agents to system. Due to asynchronous nature of system, the problem can change after initial setup and unlike some previous techniques Max-Sum can handle change in problem setting on fly and does not require re-computation from scratch.

Similar to previous work the whole framework is designed using ROS and OpenScenegraph. Besides extensive simulation, this formulation was tested using 3 mobile robots(1 Pioneer P3DX, 2 Turtlebots).

Currently this work in under review for ICRA,2014.

Floor segmentation using a monocular camera

This work is part of my colleague Suryansh's thesis and I made some initial contribution in it.

This project involves detection and segmentation of floor area using a monocular camera mounted on a mobile robot. Using different multi view geometry techniques and a color based segmentation algorithm we were able to segment out floor from a sequence of images.

This work was published in ICVGIP,2012 for oral presentation.

Exploration using a Micro Aerial Vehicle

Currently I am working on exploration of indoor environment using a monocular camera. Camera can be mounted on either a UGV or MAV. To make map I use PTAM. It is of the famous VSLAM techniques.

Due to following reasons this problem is non-trivial:

  • Pointcloud provided by PTAM is very sparse and cannot be used directly for navigation or obstacle avoidance.
  • Map is build on an arbitrary scale.
  • Tracking in PTAM goes haywire in case of rapid rotational motion, raising questions about loop closure.

The objective is to densify the sparse map information by extracting planes from pointcloud and use that for getting some sort of 3D reconstruction. For scale, odomtery reading from robot can be used to estimate the motion of robot between two initial keyframes. To prevent PTAM for breaking we plan to optimize the trajectory of robot to ensure the rotational motion is minimum and some anchor points are always visible. This video shows some primitive results of PTAM using Ar Drone.