Curriculum Vitae

Swagat Kumar
Control Laboratory 213, Western Laboratory Department of Electrical Engineering Indian Institute of Technology Kanpur Uttar Pradesh India 208 016 Phone: +91 512 259 7854 Mobile: +91 99 350 66 2 77 Fax: +91 512 259 0063 Email: [email protected] Homepage: http://home.iitk.ac.in/ swagatk/

Education
Degree Ph. D. M. Tech. B. E. 10+2 10 Specialisation Electrical Engineering Control System Engineering Electrical Engineering PCMB Institute/Board IIT Kanpur IIT Kanpur NOU, Orissa CBSE CBSE Year 2009 2004 2001 1997 1995 Scores CPI = 8.0/10.0 CPI = 9.56/10.0 80.0 % 84.4 % 77.2 %

Areas of Interest/expertise
Neural Networks Robotics - Dynamics and control Visual Motor Coordination Adaptive Critic based optimal control Nonlinear systems

Research
PhD Thesis: Kinematic Control of Redundant Manipulators using Neural Networks
July 2004 - December 2008 Guiding a robot manipulator using visual feedback is known as visual motor coordination. It involves solving inverse kinematic relationship between robot end-effector position and joint angle vector which is highly nonlinear and closed form solutions are not available. The problem becomes more challenging owing to the presence of redundancy where one may have several joint angle combinations for the same end-effector position. While vision provides very useful feedback for closed loop control, it suffers from various drawbacks like selection of suitable features, singularity of interaction matrix, pose estimation etc. Among various available approaches, learning based methods have gained importance in the recent time owing to their ability to adapt to a new system or environment. It also promises to make the scheme model independent thereby by avoiding the exact computation of model parameters. My work specically focusses on learning based strategies for solving inverse kinematic problem of a 7 DOF redundant manipulator in a visual-motor coordination paradigm. The redundancy is resolved at position level. Main contributions in this area are as follows: 1. Sub-clustering in conguration space: : A KSOM-based new network architecture is proposed that preserves the redundancy available in the training data. In this approach the input Cartesian space and Output joint angle space are discretized using a KSOM (Kohonens Self Organizing map) and each input cluster is associated with multiple joint angle clusters. It is shown that this approach provides multiple solution for a given target point and one can choose one depending on various criteria. 2. Least-square based function-decomposition scheme for pick-and-place application: The standard KSOM based approach takes multiple steps to attain a reasonable positioning accuracy which might not be suitable for a pick-and-place application where one expects reasonable accuracy in just one step. We show that the Jacobian matrices can be learnt directly from input-output data generated using a model through least square

Curriculum Vitae

techniques. However the accuracy of this approach deteriorates as the number of degrees of freedom of the manipulator increases. A function-decomposition based scheme is proposed where the effects of last 3 joint angles are decoupled from rst 3 joint angles. This decomposition leads to signicant improvement in the positioning accuracy attained in a single-step movement. 3. Solving inverse kinematics using a KSOM-RBF network through forward-inverse learning scheme: It has been shown in the literature that by inverting a forward-trained RBF network, it is possible to obtain a joint angle conguration for a given target point. However such an approach suffers from three main problems: rst, in order to ensure that the angles lie with their physical limits, one needs to provide a suitable initial guess; secondly, in order to get accurate tracking by the actual robot manipulator, the forward training requires a huge number of data sets; thirdly how to resolve redundancy for a given target point. We solve the rst and third problem by obtaining the initial guess from a KSOM-SC architecture as explained above. The clustering process can be carried out during the forward learning of RBF network itself. Since the sub-clustering architecture preserves multiple solutions for a given target point, it is possible to select a suitable initial conguration based on some criteria, like lazy-arm motion, or minimum-angle norm etc. In this way, the redundancy can be resolved directly at position level without resorting to complicated optimization schemes as available in the literature. The second problem is addressed by carrying out a forward training process within the inversion loop whenever the positioning accuracy is not met. Note that, in this forward training, a new data pair is generated using the output of the inversion algorithm. 4. Real-time implementation: All the proposed schemes are implement in real-time on a 7 DOF robot manipulator obtained from Amtec Robotics. Various experimental videos can be obtained on request.

Research visit to ISRC, University of Ulster, UK

01 August 2008 - 31 December 2008 I was a visiting researcher for a period of 5 months to Intelligent Systems Research Center (ISRC) at University of Ulster, UK. My work involved controlling manipulators using visual feedback and I implemented few visual motor coordination algorithms on mobile manipulator platforms comprising of PowerCube manipulator mounted on Scitos G5 mobile base. During this work, I acquired hands on experience on Scitos platforms and Player/Stage software.

Internship at Indian Institute of Science, Bangalore, India

December 2007 - March 2008 Optimal control for linear systems is very well established and well understood. However, the same is not true for control of general nonlinear systems. A typical nonlinear optimal control problem necessitates solving HamiltonJacobi-Bellman (HJB) equation which is a nonlinear partial differential equation. It is not possible to nd a closed form solution to this problem for a general nonlinear system due to the curse of dimensionality. Most of the time, this problem is solved in discrete-time domain as a multistage decision making problem known as Dynamic Programming. Adaptive Critic is a model inspired by reinforcement learning to solve the Bellmans dynamic programming in discrete-time. We developed a continuous-time version of adaptive critic where only one network is used to solve HJB equation in real-time so that the closed loop system remains stable during the evolution of states. As time proceeds, the controller attains its optimal value.

M. Tech. Thesis: System Identication and Control using Neural Networks

July 2002 - June 2004 This work contains two parts. In the rst part, a variant of recurrent neural network called Memory neural network (MNN) was used to learn the dynamics of a nonlinear system. Three algorithms namely BPTT, RTRL and EKF were used for training the network and a comparative analysis was carried out. A new algorithm based on Lyapunov method was proposed to speed up the training of feed-forward neural network and its performance was compared with conventional techniques. In the second part, several well known neural network based techniques were applied to control robot manipulators and comparative analysis was carried out. A new neural network based control scheme was also proposed for controlling a pendubot, an under-actuated system.

Curriculum Vitae

Publications
Journals
On adaptive learning rate that guarantees convergence in feed-forward neural networks, Laxmidhar Behera, Swagat Kumar and Awhan Patnaik, In IEEE Transactions on Neural Networks (Vol. 17, No. 5), 2006. Visual Motor Control of a 7 DOF Robot Manipulator Using Function Decomposition and Sub-clustering in conguration space, Swagat Kumar, Naman Patel and Laxmidhar Behera. Neural Processing Letters, Vol. 28, Number 1, August 2008. Visual motor coordination of a 7 DOF Manipulator using Redundancy Preserving Learning Networks, Swagat Kumar, Prem Kumar P., Ashish Dutta and Laxmidhar Behera. Robotica (Accepted) Kinematic control of a redundant manipulator using inverse-forward adaptive scheme with a KSOM based hint generator, Swagat Kumar, Laxmidhar Behera and Martin McGinnity. Under Review.

Peer Reviewed International Conferences

Kinematic control of a redundant manipulator using inverse-forward adaptive scheme, Swagat Kumar, Laxmidhar Behera and Martin McGinnity. Irish Signals and Systems Conference, Dublin, June 10-11, 2009. Implementation of a Neural Network Based Visual Motor Control Algorithm for a 7 Dof Redundant Manipulator, Swagat Kumar and Laxmidhar Behera, IEEE World Congress on Computational Intelligence, Hong Kong June 1-6, 2008. Continuous-time single network adaptive critic for regulator design of nonlinear control afne systems, Swagat Kumar, Radhakant Padhi and Laxmidhar Behera, IFAC symposium, Seoul, June 2008. A model free redundancy resolution technique for visual motor coordination of a 6 DOF robot manipulator, Swagat Kumar, Amit Shukla, Ashish Dutta and Laxmidhar Behera, International Symposium on Intelligent Control (ISIC), Singapore, October 1-3, 2007. Direct Adaptive Control using single network adaptive critic, Swagat Kumar, Radhakant Padhi and Laxmidhar Behera, IEEE Conference on Systems of System Engineering (SoSE), San Antonio, April 16-18, 2007. A novel learning algorithm for feed-forward networks using Lyapunov function approach, Laxmidhar Behera, Swagat Kumar and Awhan Patnaik, Proceedings International Conference on Intelligent Sensors and Information Processing, ICISIP, Chennai, 2004. Identication of non-linear dynamical systems using recurrent neural networks, Laxmidhar Behera, Swagat Kumar and Subhas Das, IEEE TENCON, Bangalore, 2003.

Skills
Software
Languages: C/C++, Fortran77 Operating Systems: GNU/LINUX, Windows Scientic/Numerical Packages: Matlab, Scilab, Octave, Maxima, Mupad, LAPACK, BLAS Application specic tools/protocols: Remote procedur calls (RPC), Posix threads Image processing: OpenCV Documentation: Latex, Pstricks Languages for Web : HTML, CGI, PERL Development: Developed a Neural Network Toolbox using C++ and is available for download at http:// home.iitk.ac.in/ swagatk/research.html. The toolbox presently provides various functionality related to feed-forward network (MLP) and radial basis function (RBFN).

Curriculum Vitae

Hardware
PowerCube 7 DOF Robot Manipulator (Amtec Robotics): Integration with vision system for visual motor coordination experiment. IRB 140 M2000 6 DOF Robot Manipulator (ABB Robotics) : RAPID language programming, RPC communication NI DAQ : 6025E (E-Series), 6221 (M-Series) : Data acquisition using C APIs for real-time control system experiments. Patrolbot, A mobile robot platform from Activmedia Robotics: Development of vision based tracking algorithm, Fuzzy based depth estimation and determination of angle of rotation, Real-time implementation using OpenCV, ARIA and pthreads. SGI Prism Workstation with 4 node 2 Itanium Processors, running on RHEL 3.0 : Programming and maintenance. ADSP 2101: Development of Neural network programs on DSP processor using assembly language (carried out as a part of course work for EE678) Quanser Mechatronics kit with TMSC6711 DSK: Developed several algorithms using code composure studio and implemented various real-time experiments using C/C++ and SIMULINKs Real-time toolbox.

Work Experience
Senior Project Associate at Control Lab, IIT Kanpur: Jan 2009 till date. My work involves Guiding Masters students. Carrying out research on visual servoing and Adaptive Critic Tutor for course EE250 (Control System Analysis) : Spring semester, 2008 at EE Deptt. Tutor for course EE380 (Control System Lab) : Fall semesters, 2006-2007 at EE Deptt. Teaching Assistant for Control Lab Development : Spring semesters, 2005-2007 at EE Deptt. Developed several real-time control experiments including Ball-beam System, Cart-Pole Problem. This included real time data acquisition using NI DAQ cards and interfacing them with Visual C++. I also built simulink programs for real time control of dc motors on Quansers mechatronics kit that employes Texas Instruments TMSC6711 DSK Kit. The videos of some of the experiments are available at http://home.iitk.ac.in/ swagatk/rtc_sk.html. Teaching Assistant at Advanced Center for Material Science (ACMS), IIT Kanpur : Fall 2004 Developed a web-based online equipment issuing portal using HTML, CGI and PERL.

Other Credentials
Graduate Aptitude Test Examination (GATE) - All India Rank (AIR) - 62 with 99.48 percentile, 2002 Graduate Record Examination (GRE) - 1290/1600 and Writing Assessment - 5.5, December 2003