Research

I find modern computer vision algorithms quite fascinating. I also like to deploy algorithms in robotics due to its growing applications in the society and so I work in the intersection of computer vision and robotics. I have spent some excellent summers previosly with professors in autonomy and robot perception. I aspire to contribute to research in algorithms that are trustworthy and addresses dependability and reliability in AI-driven robotics. Overall, I like to study and use different mathematical tools in complex end-to-end pipelines to gaurantee failure-resistant operation

My research draws from Robotics and Machine Learning, with the following themes:

3D Scene Understading: Inspired by the SPARK Lab at MIT, I am fascinated to work in robot perception for scene understanding and reasoning object association over time. I have worked in Long-Term object tracking in an houshold environment in collaboration with Amazon Lab126 to track and search multiple objects by leveraging contextual information to mitigate poor clasification / segmentation and object localisation issues when deployed in real-time!

Reliable Risk-Aware Predicition in Navigation: Inspired by Davide Scaramuzza from RPG group and Sebastian Scherer from AirLab, I am to leverage rich information from vision data and combine with navigation schemes for a reliable and safe planning so that industries can use it blindly in day2day life. In this regard, I worked in 2D SLAM and implemented on ground and aerial robots. However, since vision-based methods also can be fooled by symmetries in environment/ motion blur/ and high frequency demand, I like to enforce “safety” using risk-aware / degeneracy prediction methods including observability analysis to define a control input from an “allowed” set as satisfied by the constraints.

Deep Reinforcement Learning X Control: Inspired by the world and DeepSeek team, I am generally happy to use reinforcement learning and analysing sim2real possibilities. I use it for decision making to take actions according to a particular task-coded reward function. I like to research in such particular reward functions that can gaurantee agent-actions in a “safe” set as determined by the environment. The resemblance of RL reward function generation with control laws fascinates me and its ability to be generalizable over dynamics makes it cool!

Google Scholar