Automatic Reward Densification
Classical planning over human-specified PDDL models, combined with potential-based shaping, densifies sparse rewards and speeds up learning across robotic tasks.
Projects
Smart Glasses
Prototype wearable combines OCR and face recognition to support blind users with reading and social identification.
Catching a Ping Pong Ball with an iiwa
Drake-based finite-state control splits the catch into pre- and post-contact modes, combining projectile modeling with stabilized IK to stop a ping pong ball on an iiwa paddle.
Sample-Based Planning under Discrete Space
dRRT couples discrete sampling with search-based refinement, reusing computation to accelerate 5-DOF arm planning versus traditional RRT.
Lossy Compression using Neural Networks
Tested quantization-aware autoencoders that couple commit loss and training-time discretization to achieve compact latent codes for images and text.
Autonomous Driving using CNNs
Collected simulator trajectories, trained a CNN to map images to steering commands, and demonstrated closed-loop driving in the Unreal Blocks environment.
Classification for Research Universities in India
Designed quantitative criteria for research HEIs in India, applied them to top institutions, and surfaced 40 universities and 32 engineering colleges as research-focused.
Parallel Depth First Search for Directed Acyclic Graphs
Work-efficient GPU traversal expands frontier vertices in parallel and accelerates DFS-based analytics on massive DAGs.
COTTON: A Light-Weight Work-Stealing Runtime
COTTON couples work stealing with CPU frequency scaling policies tuned by task heuristics, cutting energy use while preserving throughput.
GitHub Recommender System
Collected contribution signals, formed confidence-weighted implicit ratings, and trained a custom autoencoder that surfaces relevant GitHub repos for contributors.
You can find my other projects at my my old website.