My broad research interest is building autonomous agents which can continually learn throughout its life time. I am interested in understanding, improving and implementing two aims of life long learning, knowledge representation and selective transfer. In order to achieve above goals, I primarily focus on how Robots can make use of Spatial and Temporal Abstraction to represent knowledge and reuse it whenever necessary. Robot can obtain these abstraction either autonomously or via Human-Robot interaction. I am also interested in investigating if there exists a way in which Robot and Human both can understand these abstract representations and communicate and collaborate to achieve certain goal or work on same task.
(Please visit “Projects” page for complete list of all projects I’ve done with small description links to respective Git-Repos and demonstration videos. They’re FUN!)
MSc. Thesis: Active Preference Learning Using Trajectory Segmentation
Machine learning is currently used heavily to develop robot behaviors, giving great flexibility and power, but there remains significant burden on human designers to specify reward
functions or label data for the learning target. Learning from demonstration is a very intuitive way to teach your robot a new skill. This thesis considers two improvements to
modern learning from demonstration
First, we consider a model-based imitation approach that utilizes a modern form of deep probabilistic model to predict agent behaviors in order to match them to demonstrations.
We replace the non-parametric estimator utilized by an existing approach, which has the limitation of poor scaling with the amount of training data. In its place, a learned parametric
model is trained to capture inherent uncertainties. Through sampling-based prediction, our approach is able to capture a distribution over likely outcomes of the given policy. This is
paired with a probabilistic notion of the difference between the agent’s outcome distribution and the distribution of demonstrations, to produce a gradient-based policy improvement
approach. Our results show that this method is effective in imitating demonstrations in a range of scenarios.
The second portion of this thesis considers the temporal credit assignment problem within learning from demonstration. We propose an active learning framework that uses
trajectory segmentation to addresses this issue. Our method uses spatio-temporal criteria to segment the trajectory. These criteria can be based upon speed, heading, or curve of the
trajectory which are intuitive properties to understand user’s intentions. Thus, not only does our framework make the user query interface more intuitive but the resulting approach also learns faster. We demonstrate and evaluate our approach by learning a reward function for various driving scenarios and show that our algorithm converges faster.
RJ: Radiologist Junior
During my Insight AI Fellowship I worked on developing a Visual Question Answering system for Radiology Images. Given an image and a question in natural language, the application
reasoned over visual elements of the image to answer the question in natural language.
Medical Assistant Arm: MAA
Team: Bartosz Miselis, Justin Whatley, Nima Amooye Foomany, Francis McEachern and Monica Patel.
During my masters I was also enrolled in a Surgical Innovation Graduate Certificate Program offered by Experimental Surgery at McGill. Throughout this program we went through Need Scoping exercise in field of medical technology. We were exposed to current working medical environment and we worked in multi disciplinary team to develop working prototype of novel medical technology. For this project I worked on building a medical assistive robot arm that will recognize, count and manipulate tools in operating room in order to help scrub nurses better assist surgeons. The prototype application that recognized and counted tools was developed using Kinova JACO2 manipulator arm. Process of building the data set of surgical tools was also automated using the same arm.
Curio: Educational Robotic Platform.
Team: Abhishek Kulkarni, Monica Patel, Shruti Phadke and Aditya Joshi.
As my undergraduate BTech. Project I worked on building a Low-Cost High-Tech Educational Robot which can be used as learning tool by students and researchers and as teaching aid by faculty members. One of the salient feature of the robot is its software library which is based on subsumption architecture and uses extended state machine model to run each block. This makes the library modular to use in educational setting. That is, a student working on Robotic Vision will not have to worry about Robotic Control or Localisation and vice versa. We also programmed a set of Python libraries that provide easy access to popular algorithms and basic robot control behaviors. These behaviors included moving robot in certain shapes, localizing robot in indoor environment with over head camera and building approximate map of surrounding using ultra sound sensors.