Application for MS in Robotics at EPFL

Name: Sai Siddartha Maram | Email ID: ​[email protected]

“Congratulations : ​Research Paper Accepted (Robotics, Computer Vision) at ​IEEE ICACCP 2019”
was the email which reflected that walking the extra mile for ​research​, beyond the realm of my
curriculum and interning at Invento Robotics was fruitful. Having a paper accepted at IEEE ICACCP
which has a mere ​acceptance of 26% motivated me to further work on 3D Perception. This work helped
me bag the position of ​Research Intern at ​Georgia Institute of Technology​, Atlanta. I also filed a
patent for an IoT and Computer Vision prototype for the problem of physical abuse. This illustrates my
passion for pursuing research to create a global impact and this resonates with EPFL’s vision of being a
research university to create impact. I would like to delve deeper in the field of Robotics for which I am
seeking admission at EPFL’s reputed Robotics Graduate program.

During the summer of 2016 I interned at SRISTI-UNICEF under Dr.Anil Kumar Gupta at the ​National
Innovation Foundation, Government of India​. SRISTI allowed me to connect with farmers and discuss
problems they faced while farming. ​Empathizing with farmers​, we noticed an intricate problem they
faced which is animals vandalizing crops. We started exploring Computer Vision solutions to deal with
this problem. To start off my team ​implemented the Paul Viola and Michel Jones paper ‘Rapid Object
Detection using a Boosted Cascade of Simple Features’ to develop a Haar Cascade suitable for detecting
cattle. Tackling detection problems through this algorithm which took only edge and pixel intensities as
features were less fruitful as even the slightest change in orientation of cattle proved to be a failure. We
noticed we needed to incorporate more features for detection, for which we applied transfer learning on
the Faster_RCNN_Resnet model trained initially on the COCO dataset. After training and comparing
results over tuned hyperparameters, we ​published our results at TEAMC 2018, an undergraduate research
paper conference, where we ​won the​ ​Best Research Paper.

A good grip on Data Structures and Algorithms along with an inclination towards research in Computer
Vision has helped me get selected as a ​Summer Intern by Mr. ​Balaji Vishwanathan (​CEO Invento
Robotics​) for my summer break 2017. This internship allowed me to get a first-hand experience with 3D
mapping using LiDAR and ROS (Robot Operating System). The robots limited hardware did not allow
the deployment of dense neural networks and hence initial developers had used Cloud services for object
detection. This stack was highly dependent on internet bandwidth was time-consuming. I was able to
achieve a ​60% faster ​performance than the previously established stack by choosing to train and ​deploy
a less dense neural network architecture ‘tiny-yolo’ for detection making it suitable for embedded systems
on Invento’s flagship robot ‘Mitra’ (India’s first humanoid). ‘Mitra’ is currently used for navigation by
travelers at the Indira Gandhi International Airport, New Delhi, India and uses detection algorithms that I
had developed during this internship.

My experience with Computer Vision algorithms on ROS at Invento paved the path for my next ​research
paper​. In this paper, we discussed how Neural Networks could be treated as ROS nodes and keep
publishing outputs in the form a ROS message as a stream to the ROS base node which is used for
navigation and motion. After training the Neural Network (for threat detection) node on various
hyperparameters, we ​published ​the performance of our algorithm at ​IEEE ICACCP 2019 ​where we ​won
the Best Research Paper ​under the computer vision and robotics category.
I lead my team into solving the problem of physical abuse against women. We developed a cognitive
textile that used Computer Vision for the ​first time to tackle such abuse against women by analyzing live
camera feed through cameras embedded in it. This IoT and AI prototype has been filed for an Indian
Patent (​A personal safety device and method thereof​) ​bearing application number 201911005811. This
patent is a reflection of my vision of amalgamating technology and research for solving real-world
problems which is a primary objective of research undertaken in EPFL.

I shared my paper at IEEE ICACCP 2019 with Prof. Yi-Chang Tsai, who then offered me a ​Summer
Research Internship at ​GeorgiaTech, Atlanta​. During this internship, I applied the concept of
perspective transform suitable for United States interstate roads for traffic sign depth estimation through a
single 2D image​. I was able to ​reduce memory consumption ​by 50% than the regularly used
triangulation method which requires two images. One of my novel ​research outcome as a result of this
internship was that I managed to develop a 2D image and 3D LiDAR point cloud registration algorithm
by using a ​data structure ​called KD-Tree​. This leads to enhanced performance compared to existing
work on huge LiDAR point clouds. By thresholding the LiDAR point clouds retro intensity values, I was
able to accurately estimate the health of all traffic signs over the entire length of any interstate.
Implementing clustering techniques such as HDBSCAN for GPS points instead of the conventional
K-Means and DBSCAN allowed to eliminate computations on bogus LiDAR points making runtime
quicker. Modern robots use LiDAR for navigation purposes, I am curious on how I would be able to bring
research experience at GeorgiaTech on LiDAR into EPFL’s reputed LASA Laboratory.

I will be interning as a ​Research Assistant at the prestigious ​Indian Institute of Science, Bangalore
(Spring 2020) under ​Prof. Yogesh Simmhan dealing with AI for drones. EPFL is a place where research
meets empathy and compassion and it has impacted millions of lives around the globe. An admit to EPFL
will give me an opportunity of working with a diverse peer group under the able guidance of elite
professors and undertake research in Robotics to develop solutions for global problems. I see myself as a
part of research labs or academia to pursue research for overcoming challenges through technology and
ensure that research results are not confined to journals and conferences but reach masses.