• Aishwarya Pothula

Spring 2020 Semester Review

Updated: Sep 15, 2021

The plan this semester regarding SEDRo was to start versioning of our implementations as the basic environment is ready. We are to focus on implementing vision and touch sensors for the baby agent and also an experiment from the field of developmental robotics - Occluded rod display. I also plan to build a table of differences between five relevant papers for our research and our work. This will help me in developing the related work section in papers to be submitted.

I plan to read the Touchdown and Foundations for New Science of Learning papers and summarize them. In terms of Udemy courses, I plan to complete the "Ultimate Guide to Game Development in Unity" course and the "Reinforcement learning in Python" course. Additionally, I will be working on solving the problem in the PRML book and writing the solutions in Latex to publish them.

This past few months have been strange, taxing and anxiety-ridden , not only for me but for the entirety of the human race thanks to COVID-19. I started the semester with a trip to India for Visa work. I came back to 1 week of catching up of classes and having to prepare for a talk. I took part in the SCRF event conducted by UTA's CSE department. I was awarded the first prize for my lightening talk about about the research work at HDILab. I understood that you don't have to include every detail when talking about your research. My first trial run in the lab (thankfully, I did one) was a disaster; the room fell asleep. I made the presentation to a resemble a reasoning class; it was too technical and dense. Later, taking everyone's suggestions, I altered the content to be a simple overview of the research and included a few example of how we built our problem statement.

In terms of SEDRo we started by uploading the first version of SEDRo environment - a mother, a baby, 2 bhk house with some interactive play things in the infant's room. This version also contained an action sequence in which the mother character walks up to the cradle and bends over the crib to say "Hi" to the infant. In the second version of SEDRO, we have implemented central and peripheral vision to the baby agent, dynamic touch sensors to get the touch observations. We have also replaced out initial model of the baby agent with Unity's walker agent model. The baby's agents movements are very jerky and extremely fast. A python interface to interact with the environment executable through Jupyter notebook has also been developed. I have implemented the occluded rod display to be used in the Unity Perception experiment in third version. The experiment is to test whether infants are born with a capability for Unity perception. The ability to experience partly occluded objects as unified objects is called Unity perception. One of the issues was that the baby agent kept falling out of the crib despite invisible walls around the crib and collision detection. In the next version, I have created a the experiment to be launched as a separate scene. In order to build a scanplot of the agent's eye movements in IPython, we thought of adding the eye contact observations as part if the agent observations. On Dr.Park's suggestion, since we don't use these observations to train the agent, I implemented a side channel to communicate the eye contact observations to IPython. I have the code ready for building a scan plot from these observations. However, issues from the pervious versions persist along with joint detachment when the agent is moving. Once these issues are rectified, we can easily build the scanplot. At this point, we submitted a paper to ICDL-epirob without including the scanplot. Version Five is the implementation of a grabbing functionality for the agent. We also developed a basic environment for when the baby agent is in the womb. Right after spring break we also submitted the paper to ICML.

Regarding academic papers to read and summarize, I planned to read Touchdown and Foundations for A New Science of Learning. However, I could only read and summarize the later. I am yet to write summaries for two of the five papers essential for our research. I am yet to make a table of differences between our research and these five papers. Aside from this list, I have read the Virtual Rodent paper and summarized it.

Coming to self-study, I only partially completed my plan. I have started reading PRML but I have not attempted any problems. I have not started any Unity course. Learning Unity courses takes a lot of time, especially setting up Unity and there are often version mismatches with the tutorials leading to a long search to fix bugs. I got impatient and did not work on Unity this semester. I thought it was okay to delay doing the courses. However, I need to build expertise in Unity as I want to work in the industry after I graduate. It is important that I keep at it even though it takes time. There is no hurry to finish the courses, but I need to be persistent. I plan for a combination of Unity and Developmental Robotics to be my speciality. I worked on the Udemy Reinforcement Course and completed four sections of it. I should make the Unity and Reinforcement Learning courses a priority. Apart from the courses, I planned to start solving and publishing solutions to the PRML book. This task too, I started with a lot of enthusiasm and abandoned midway.

This semester was busy with Neural Networks and Reinforcement learning classes. The workload of both courses was heavy. However, taking Networks with Dr.Kamanagar was an exhilarating experience. He was just brilliant as a teacher in the way he would connect seeming disparate concepts to make us develop myriad perspectives about a concept. For example, to explain decision boundaries for classification, he spoke of something as inconceivable as the space time continuum and, yet, found a way to beautifully connect them. On the other hand, I really put in the hours to complete the reinforcement learning assignments and projects. These assignments have singularly lead to a vast improvement in my coding skills. The knowledge I acquired to complete these assignments really made me comfortable with a lot of RL concepts. I also published an article about POMDPs after doing an assignment based on them. I still maintain my 4 point GPA :) (touch wood)

Submitting a paper to the ICML-Lifelong Learning workshop as practice was a big learning experience for me. I made many mistakes and I'll be sure not make them again. First thing, I started late, I really underestimated the time it would take me to write the paper. I was under the impression that since I already have paper that we have written (or rather edited - Dr.Park wrote the papers) for NSF and ICDL-epirob, it would be fairly easy for me to finish this paper. And oh boy, how wrong was I ! Though I understood the content in the paper, I really struggled to put it in my own words. Second mistake was to use a few of my sentences and image descriptions from the previously submitted papers. I have come to know that it falls under self-plagiarism. The icing on the cake was that I finally submitted the wrong paper to the workshop and spent the rest of my week trying to retract my submission.

I again travelled to India during Spring Break to attend a wedding and have been here since because of the country's COVID-19 travel ban. In retrospect, I kind of relaxed too much academically after coming to India. I should learn to be more consistent and disciplined.

I enjoy and love my work at the lab and would benefit and learn by continuing to work in the lab. I sometimes experience short phases when I don't feel like working. (I guess that's okay!). The greatest challenge this semester, apart from my usual hard pressed for time, was to be focussed. I kind of lost track of the things to be achieved this semester and fell behind on my plan for the semester. Dr.Park has been very supportive and encouraging throughout the semester.

I am trying to follow the lab culture to the best of my abilities. I try to collaborate with my lab mates whenever possible. I am trying to create reproducible output. I am earnest in my work. I have the drive to learn and excel.

My plan for the coming semester is to make significant progress in implementing experiments from developmental robotics such as the occluded rod display experiment and write a paper "Automatic Evaluation of (VExP) Developmental Progress of Non-verbal Agents".

On the self-study front, I plan to attend the Neuromatch academy workshop, solve 1 problem per week from the PRML book, work on the Unity and RL Udemy courses, and also on the tasks remaining from this semester's plan.Additionally, I want to enroll in an academic writing course.

1 view0 comments

Recent Posts

See All