Research Statement & Interests

Areas of Interest

I. Automation and Robotics for Human-centered Environments

II. Automation and Robotics for (Hazardous) Fields

III. Machine Learning for Robots

My research interests lie in building intelligent robots for meaningful purposes - particularly, service / domestic / assistive robotics.

I am keenly interested in how knowledge can be conveyed to robots and how high-level knowledge and reasoning capabilities of robots are connected to the way robots view and understand the world - through sensors, vision systems, or haptics.

I. Automation and Robotics for Human-centered Environments

a. What kind of robots will I work on?

From my time at the University of South Florida (USF), I have been introduced to the ever-growing field of robotics. This area is large, ranging from problems such as building and designing robots to using robots for various applications (viz. medicine, navigation, transportation, industry, etc.). For me, I am particularly interested in robotics applications that can help people who are impaired in some way through service robotics - in particular, domestic robots for the household that will interact with humans. In addition, service robotics will also improve humanity's quality of life - we can develop robots that can work in place of humans in dangerous conditions or environments. , I also want to venture into developing robots that can help the impaired (the elderly or physically disabled individuals). In addition to these kinds of robots, although I am not currently experienced in this multi-disciplinary field, I am also interested in other applications of socially assistive robots (SAR) such as rehabilitation for mental disorders or physical ailments. I would love to take students on that have any interesting ideas for applications that have direct impact on these target groups and more.

b. Knowledge Acquisition, Representation, and Understanding

I am interested in how we can teach robots how to perform tasks and how we can have robots that behave and think in a high-level manner. A robot that can acquire knowledge and use it to perform tasks intelligently will be of more use than robots that are hard-coded to perform some task repetitively. I believe that intelligent robots can enforce safely and reliability, while building upon the principle of explainable AI. From my work at the Robot Perception and Action Lab at USF, I extensively developed a knowledge representation for task representation and planning called the functional object-oriented network (FOON). More details on this representation can be found here. Through this work, I learned a lot about graph theory and different considerations needed to effectively build representations of knowledge: we made FOON as a bipartite network to capture the relationship between objects and motion types as two separate types of nodes. Graphs have the ability to conveniently convey information in a way that is easily read by humans. We have developed algorithms for retrieving knowledge from this network.

However, this is still far from becoming a standalone representation, and what I would like to do is do more research on the ideal ways of developing knowledge representations for robots. There are still important questions and problems to consider to build such systems. Along with high-level reasoning capabilities and semantic knowledge, we must not forget the importance of low-level motor skills and state recognition / representation. Oftentimes, researchers do not consider these two steps as interconnected problems and they are usually solved independently. I aim to study how high-level states (such as those we considered in FOON) can be related to how the robot perceives its environment and task space - typically using sensors, cameras, and haptics.

c. The Importance of Artificial Intelligence

The way human minds work (to how we achieve individual thoughts, desires, beliefs and actions) has always been a topic of interest. This is one of the main motivators of artificial intelligence: to answer how our cognition and sense of "self" arises, and introspectively understanding how we can artificially create a being with similar sentience. I believe that artificial intelligence goes hand-in-hand with robotics, because in order to build an intelligent robot, we should build it with intelligent behaviour. For instance, we can interact with robots for feedback on the course of actions they take (which is now an ongoing field of explainable artificial intelligence) or in what ways they can perform their actions better. Explainability is also a very important problem to be solved in other AI-based applications, where there are obvious biases that are present in decision making algorithms or recognition models.

II. Automation and Robotics for (Hazardous) Fields

In addition to service robotics, I wish to explore the automation of certain processes or tasks needed in areas such as agriculture, mining, manufacturing and other important industrial fields. Instead of seeing automation as a bane to human employment, I want to explore how robots and humans can work side-by-side towards a common goal. This would involve identifying key parts of these fields that can be performed by a robot, especially tasks that are very unsafe for humans. This may not require complex robots and can involve unmanned aerial vehicles (UAV) or simpler mobile robots, incorporating important tasks such as navigation and SLAM. In particular, I would like to make a case study of a certain population or group to identify the risks involved in a particular field, the improvements (if any) that can be made to performance and throughput by introducing robots, and the trade-offs of using robots in these processes. In certain instances, it may be necessary to customize or fabricate robots that will be simpler yet more affordable for impoverished groups or societies, which would require collaboration with experts in other disciplines and areas of focus.

III. Machine Learning for Robots

Finally, I am interested in solving or improving minute aspects to robot learning that can be learned through machine learning and reinforcement learning. These techniques are ideal for learning policies or patterns that a robot can use as cues for object detection, instance identification, and image segmentation. Embedding words and concepts for robots is also a relatively new problem that is important for accurate representation of concepts that are learned in an unsupervised manner. I have proposed a new method to embed the concept of motions or manipulations in a way that would make sense to a robot's operation or understanding of action, which we coined as the motion taxonomy. This work, although fairly new, shows promise in learning embedding that can improve the performance of motion classifiers and activity recognition models.

However, there remain several problems that have not been clarified in robotics using such machine learning techniques. For starters, although neural networks are excellent for classification tasks of what they have been trained to do, they are not great at handling cases in which they have not been trained for. As mentioned in my survey paper on knowledge representations in robotics, this relates to the problem of appropriately describing the problem statement and handling the issue of open set recognition. This severely limits the potential of deep learning and machine learning to the real world.

I would also like to explore how we can improve the performance of these models as they relate to robotics - specifically through some theoretical contributions.

A take-away from a keynote speech given by Dr. Wayne Book from Georgia Tech (given during IROS 2016) for me was this: we should use our talents to helping those in need. During his speech, he mentioned his efforts to help the impoverished in Haiti. I too aim to use my research and talents to solve problems for communities or groups in need, where robotic solutions can greatly benefit a society and where robots can efficiently and safely solve important day-to-day problems.