Automatic Reward Densification
We implemented a system that is able to leverage classical planning over humanspecified PDDL models to automatically increase the density of robotic tasks with sparse, goal-based reward
Projects
Smart Glasses
A prototype to help blind people understand text in books and also allow them to identify people they already know in order to improve their life style. It was selected to be showcased at Delhi Mini Maker Faire.
Catching a Ping Pong Ball with an iiwa
We programmed the kinematics of a Kuka iiwa using Drake to catch a ping pong ball with a ping pong paddle.
Sample-Based Planning under Discrete Space
Proposed a hierarchical decomposition algorithm where we discretize the continuous sample space of PRM/RRT algorithms in order to provide tighter completeness guarantees. Demonstrated the performance improvement of the proposed approach for 6-link robotic arm
Lossy Compression using Neural Networks
Formulated quantization techniques to generate discrete latent space representations among image and textbased autoencoder models without significant performance implications. Showcased that incorporating commit-loss to the learning process improved the compression ratio of both imageand text based models while maintaining the quality of reconstructions
Autonomous Driving using CNNs
Trained a custom CNN model on simulated data generated form Unreal Engines’ Blocks environment. Demonstrated the efficacy of the model to steer the environment autonomously via control inputs given by the model.
Classification for Research Universities in India
Worked on an extended classification approach for research universities in India, based on the Carnegie classification approach. Proposed a simple basic criterion for identifying research universities depending on factors such as research funding, number of faculty members etc.
Parallel Depth First Search for Directed Acyclic Graphs
Implemented a parallel version of the popular Depth First Search algorithm which is by nature non-serializable, written entirely in CUDA and showed significant performance improvement over the serial version over large graphs.
COTTON: A light-weight work-stealing runtime for async-finish parallelism
Developed a light-weight work-stealing runtime for async-finish parallelism which was energy efficient without incurring significant impact on the performance. Used different power saving drivers in combination with cpufreq to change the CPU frequency based on some task based heuristics.
Github Recommender System
Developed a recommender system similar to Amazon/Flipkart for Github where users are provided with good repositories for working on. Aim is to promote open source culture. Obtained a recall score of 0.72
You can find my other projects at my my old website.