Robotics Researcher and Engineer
I am an MS student at the University of Washington, Seattle, specializing in robotics research and engineering with a focus on computer vision and controls of robots. Currently, I hold the position of Research Assistant at the Robot Learning Lab, working with Prof. Byron Boots.
Interests
Vision-based controls and navigation
Self-supervised scene understanding
Whole-body control
Learning-based control
Full Stack Robotics
Background
Hello there! Thanks for taking the time to visit my website. I hail from Navi Mumbai, India, and now call Seattle my home as I pursue my MS degree. My upbringing was blessed with supportive parents who encouraged me to explore beyond academics. Playing the piano, painting, and my passion for football became integral parts of my life. I cracked the IIT-JEE Advanced and got into Mechanical Engineering at the Indian Institute of Technology, Gandhinagar. As captain of the college soccer team, I represented the squad in state and national-level tournaments. Additionally, I was part of the college band, performing at various fests and arenas. I was awarded the Institute's Silver Medal and Director's Silver Medal for outstanding performance in academics and all-round performance respectively.
Research Journey
My research journey started at the Indian Institute of Technology, Gandhinagar (IIT Gn),
with building an autonomous 1/10th scale race car to control at high speeds. We used a novel
temporal shallow-deep neural network to detect roads, a genetic algorithm for local planning,
and PID to control the car. This helped me gain insights into the interplay of perception,
planning, and control and led to my first publication "Memory Guided Road Detection".
Concurrently, my involvement in the All-Terrain-Vehicle (ATV) building team for off-road
terrains made me question the applicability of our on-road algorithms in off-road settings.
This channelized my interest in robotics for unstructured environments.
At the University of Washington, Seattle (UW), under Prof. Byron Boots, I have been working in
the perception team of DARPA RACER, which focuses on running ATVs in unstructured off-road
terrains. Our team uses point-cloud data for semantic segmentation to determine the local map
around the vehicle. Point-cloud data becomes sparse with distance, and the segmentation model
is over-confident in the sparse regions. Thus, I added an aleatoric uncertainty module to quantify
the model confidence to facilitate safer navigation. Sparser point-clouds also mean the risk of
undetected objects, prompting us to use object detection techniques using camera images. However,
their practicality is limited without post-processing steps like object tracking and temporal
aggregation, leading me to develop a real-time multi-object tracker. RACER's reliance on multiple
LiDAR and camera sensors for advanced off-road navigation raises a key question for my Ph.D.
research: “Is efficient locomotion and navigation with fewer sensors achievable?" Addressing
this could reduce costs and power consumption, making robots more practical for real-world applications.
Working under Prof. Harish PM, at IITGn, in our journal article "Impact of Added Passive Compliance on the Performance of Tip-Actuated Flexible Manipulators" taught me the practical significance of the theoretical concepts from control theory. At UW, collaborating with Prof. Guanya Shi, I was introduced to Reinforcement Learning (RL) and Optimal Control / Model Predictive Control (MPC), deepening my appreciation for controls. In our paper "DATT : Deep Adaptive Trajectory Tracking for Quadrotor Control", we developed a novel model-free RL framework for tracking infeasible trajectories, adaptable to external conditions like wind and payload, on low-cost drones, showing a 54% improvement over MPC. Next, we identified the strengths and limitations of MPC and RL. This led us to think, "Why not blend both to leverage the robustness of model-based MPC and the adaptability of model-free RL?" Consequently, in the paper, "Deep Model Predictive Optimization", we showed that blending MPC with RL enables drones to perform high-precision tasks like agile flips. This gives me the promise that learning-based control enables robust and agile navigation on low-cost robots and has brought them closer to real-world applications.
