INTERVIEW PROCESS & INSTRUCTIONS
There are two parts to PhD interview process: Online written test, and an in-person interview.
There will be an online test on May 13th, 2 – 3PM. The platform is hackerearth.com, and a link will be sent separately on the day of the test. The test will consist of two parts a) Objective questions, which will be 30 minutes long and b) A subjective question, which will also be 30 minutes long. For the subjective question, you are required to work out the solution in a sheet of paper/electronic pad and upload. The topics are foundational, typically taught in first year UG courses.
The instructions about the online test platform (hackerearth.com) shall be sent over email soon.
Note that the test scores will be used in ADDITION to the interview for shortlisting, i.e., all candidates must attend the in-person interview on May 20-21 in your allotted slot.
The slots for each applicant have been sent previously in the interview letter. The duration of the interview may vary from 15 minutes to 45 minutes. To help with the evaluation, each applicant is required to upload slides (5 max) about past projects/contests or any interesting work that she/he may have done so far in school or work.
Our questions will be based on the slides and on the online test. We will also be asking questions on foundational topics (listed below). You need to be prepared to answer questions on two out of the four topics
Linear Algebra & Matrix Theory, Calculus & Differential Equations, Discrete Math & Algorithms, Probability.
The level of questions is at the basic under graduate level.
The interview will be held at the Robert Bosch Center for Cyber-Physical Systems, located on the 3rd floor, SID Entrepreneurship Centre, IISc.
The list of PhD topics are given below:
Brief description of the projects
Security and Privacy in Networked Control Systems
(Prof Vaibhav Katewa)
The students will work in the broad area of Cyber-Physical Systems and Networked Control Systems. The project will be interdisciplinary in nature and require tools and techniques from control theory, optimization, signal processing, AI/ML etc. The project will be mathematically involved and requires good background in linear algebra, probability, calculus etc, which can be built by taking courses at IISc. We will study different types of possible attacks on network control systems, analyze their effect of the system performance, develop data-driven and learning-based attack detection algorithms, and countermeasures against such attacks. There will be experimental opportunities to implement the algorithms on drones and ground robots. For more details regarding the research and list of publications, visit my webpage – https://cps.iisc.ac.in/faculty/vaibhav/
High-Precision Control of Laser Beams
(Prof Radhakant Padhi)
Laser beams have a wide variety of applications such as simple pointers in presentations and high-bandwidth wireless communication. High-power laser beams can be used as a powerful weapon too for nullifying targets. For it to be effective, a strong requirement is that it must continuously fall on the specific location of the target for a finite amount of time (called as “dwell-time”). When the target is usually at a far-away location, the pointing accuracy requirement becomes of the order of a few microns (micro-radians). Moreover, since the beam travels through the air medium, it is subjected to a host of disturbance effects such as jittering, splitting, bending etc., which are functions of humidity, temperature, wind condition, distance of travel etc. Such disturbance effects must be rejected by the system. The mechanism that helps achieve this objective involves both coarse correction by swiveling (turning) the entire setup and fine correction by actuating a set of fast-steering mirrors. Concepts from artificial intelligence (AI) will be used to compensate both for the unknown modeling errors in the actuation system as well as for the laser beam propagation model.
(i) (Un/Semi)-supervised Learning for Big Data Analytics (ii) Graph Neural Networks for Unsupervised Learning and their applications iii) Spatio-temporal Graph Neural Transportation Analytics
(Prof Punit Rathore)
(i) (Un/Semi)-supervised Learning for Big Data Analytics: The objective of this project is to address the challenges of real-world big data such as volume, velocity, variety, and unlabeled (or partially labeled) data by developing novel, scalable and efficient learning algorithms.
(ii) Graph Neural Networks for Unsupervised Learning and their applications: Most graphs in the wild are unlabeled; however, it is still largely understudied. The objective of this project is to extend graph neural networks for unsupervised learning such as graph optimization, clustering, and graph sampling, and validate them for real-world data.
iii) Spatio-temporal Graph Neural Transportation Analytics : We will explore how spatio-temporal GNN framework can be used for transportation Optimization. More details, to be shared later.
Robust and agile design of walking robots
(Prof Shishir N Y)
Here we are mainly interested in novel designs/mechanisms that improve the overall efficiency and agility in walking robot platforms. We have a custom-built quadruped, called the Stoch, wherein we have made improvements in the mass and intertial properties of the legs. We propose to build upon these findings and formally develop methodologies that help realize dynamic locomotion behaviors like running/hopping/jumping etc.
Safety-critical control in autonomous navigation
(Prof Shishir N Y)
The project here focuses on development of real-time control laws for nonlinear dynamical systems that have guarantees of safety. Some of the examples include obstacle avoidance in drones, safe navigation in autonomous cars etc.
Learning for legged locomotion
(Prof Shishir N Y)
We are mainly interested in incorporating learning along with control for realizing robust locomotion behaviors in a custom-built four-legged robot, Stoch. We are looking for students who are interested in either development of novel learning algorithms that enable fast and robust locomotion behaviors on a diverse set of terrains.
Learning + control
(Prof Shishir N Y)
We will explore approaches for using data gathered from systems to try to control them better. Classical control techniques are heavily reliant on models of systems, which are often difficult to characterize. On the other hand, machine learning techniques rely on large amounts of data to learn a given task. What is a good ‘middle ground’ that can combine the strengths of model-based control with data-dependent learning to give algorithms that can learn to perform a task (e.g., robotic control, legged locomotion) without requiring too many data samples? We try to answer this question.
Learning-based controllers for Spatio-temporal logic specifications
(Prof Pushpak Jagtap)
Nowadays engineering systems are expected to do complex tasks such as “pick up the packet from location A and drop it at location B within 30 min and always avoid obstacles in the path and if the battery is low, go-to charging station and stay there for 2 hrs” which contains logical and temporal elements. In this project, we will explore the efficient design of reinforcement learning algorithms and control policies for performing complex temporal and logical tasks using model-free as well as model-based learning algorithms
Design and control of cooperative aerial manipulator
(Prof Pushpak Jagtap)
Aerial manipulation combines the versatility and agility of aerial platforms with the manipulation capabilities of robotic arms. However, the design and control of these aerial manipulators is a very challenging problem due to the involvement of many dynamic elements. On the other hand, the payload capacity of the aerial platforms is a major issue in the usability of the technology. This issue can be handled in a distributed manner that is by using multiple aerial platforms to handle a common heavy object. However, these required a lot of coordination and cooperation among aerial manipulators. The initial phase of the project will focus on robust stabilization and control of a single aerial manipulator and the later phase will focus on the design of cooperative and distributed control algorithms to handle common objects using multiple aerial manipulators.
Formal Guarantees for Learning-based Control of Cyber-Physical Systems
(Prof Pushpak Jagtap)
Dealing with cyber-physical systems (CPS) consisting of components of different nature described by differential equations, lookup tables, transition systems, and hybrid automata, describing the CPS in a closed-form model is often complex. In this case, a common practice is to make use of learning-based approaches, a solution that can be also motivated by the recent technological advances in sensing and processing technologies and unprecedented opportunities for collecting data at high details and at large scales for CPS. However, at the same time, learning-based components are typically viewed as black box-type systems, lacking formal guarantees. In this project, we aim at developing theoretical and algorithmic approaches for guaranteed learning-based control of CPS against complex logic specifications. To solve this issue we utilize analytical proofs from classical control theory and algorithmic proofs from the formal methods community.
Formal control of multi-agent systems against complex specifications
(Prof Pushpak Jagtap)
Multi-agent systems (MAS) under complex Spatio-temporal logic tasks have great potential due to their ability to deal with complex tasks. The control of these systems, however, poses many challenges, and the majority of existing techniques are state-space discretization based and result in a large computational burden. The research area will focus on the development of discretization-free and computationally efficient approaches to synthesize controllers for MAS with different settings (such as leader-follower MAS, MAS with heterogeneous agents, consideration of nonholonomic dynamics, and complex agent dynamics) to perform complex temporal logic tasks along with conventional objectives in MAS (such as collision avoidance, consensus, formation control, or connectivity maintenance). We will also try to address the following questions: How to decompose global specifications to the local ones to compute decentralized or distributed control laws? How can one modify an interconnection topology that guarantees the least violating solution in the presence of conflicting local complex tasks?
Multi-agent learning for coordination
(Prof Pavan Kumar Tallapragada)
Coordination between different “agents” is critical in a lot of applications. Examples include communication and control over shared channels and multi-robot coordination tasks. In this project, we will explore fundamental design challenges in multi-agent learning for coordination as well as explore some interesting applications.
|Study of transportation choice models using large scale agent based models|
(Prof Rajesh Sundaresan)
Agent-based simulators provide a way to study emerging outcomes out of behavioural choice modelling. They provide a means by which policy makers can explore various policy designs and their outcomes. We will be looking at two aspects. The first is to build a reconfigurable platform that can be model a variety of behaviours of individual agents. The second is to explore various choice models to demonstrate the benefit of collaboration, cooperation, competition, and coopetition in city-scale mobility networks.
CoBot for Healthcare
(Prof Pradipta Biswas)
This project will investigate different use cases and interaction scenarios for both collocated and remotely located Collaborative robots (CoBots) in healthcare. The research should consider subjective feedback, feasibility, user acceptance for different human-robot interaction situations involving multimodal interaction and extended reality (AR/VR/MR).
Distributed Coordination and Collaborative Decision Making for Multi-Drone Sensing and Edge Analytics
(Prof Yogesh Simmhan)
Wide Area Control of Large Power Systems with Communication Delays
Details will be shared later
This Project aims to develop cyber resilient wide area control schemes for smart grids in the presence of communication delays. The project involves development of a learning based control architecture which can cater to both the delays and cyber-attacks. The project also involves implementation of the control schemes in hardware and validation on a physical test bed.