Research Areas

The following are the current research threads pursued in the group

 

Human Robot Collaboration:

We are currently developing algorithms which would allow seamless integration of human and robot capabilities in a given task. Our focus spans across pure teleoperation to shared control to physical human integration. To this end, we are exploring ways of integrating deep learning research with classical optimal control theory based human motor control.

 

Planning and control under uncertainty:

For robots to successfully transition from structured factory floors to real world environments, they need the ability to explicitly consider uncertainty in their perception and own motion to optimally trade-off risks and cost associated with their actions. In our research group, we are looking at this problem using the template of chance-constrained optimization. The core focus is on solving chance-constrained optimization while relaxing the assumption of Gaussian distribution in the underlying uncertainty and linearity/convexity in the problem structure. We have recently proposed a computationally efficient reformulation of chance-constrained optimization under Non-parametric uncertainty and non-convex constraints.

 

Optimization for Robotics:

In recent years, there has been a strong push in the machine learning community to understand how optimization algorithms developed in the mathematics community can be best adapted to machine learning problems by exploiting the problem specific mathematical structures. However, the robotics community has not seen similar initiatives although its advantages are well acknowledged. In our research group, we are planning to fill this knowledge gap by developing optimization algorithms which exploit the structures present in robotic motion planning and control problems. To this end, we have recently shown how optimization problems involving manipulators can be reformulated to have a multi-convex structure. Several extensions to applications involving autonomous cars, quadrotors are currently being developed.

 

Machine learning based Control in safety critical applications:

Control policies computed through end-to end learning do not come up with safety guarantees that more conventional motion planning and control approaches have, although the latter works with more assumptions on the problem structure.  In our research group, we are developing ways to optimally integrate deep-learning based approaches with classical control theoretic algorithms for applications like autonomous driving and human-robot collaborative manufacturing.