Devi Aditya Varma
Dantuluri.

01

Semantic Mapping for Low-Cost Autonomous Vehicles

Standard 2D-LiDAR maps are blind to floor-level lane markings, which makes legal lane-constrained navigation impossible without expensive 3D sensors or hand-built HD vector maps. I fuse low-cost 2D LiDAR with monocular vision to generate Semantic Occupancy Grids that bake visual lane constraints directly into the navigation costmap. The pipeline uses Inverse Perspective Mapping, HSV segmentation, morphological filtering, and a temporal-accumulation layer in the ROS navigation stack.

02

Decision expected Jun 2026 · primary author

Reliability Growth Modeling of Agent-Driven GPU Kernel Optimization

LLM coding agents now iteratively optimize performance-critical software like GPU kernels, but practitioners stop iterating based on intuition rather than analysis. I treat these agent loops as software reliability growth processes, fitting five non-homogeneous Poisson process models — Goel-Okumoto, Musa-Okumoto, two Yamada S-shaped variants, and Pham-Zhang imperfect-debugging — to per-iteration AutoKernel trajectories. The output is a deployment-risk-budgeted stopping criterion: when does the expected marginal value of further iteration cross zero?

03

Decision expected Jun 2026

Physics-Informed Perception Reliability

Camera perception degrades when condensation, frost, or ice forms on the lens enclosure during temperature transitions — a short-lived but safety-critical failure that aggregate benchmark accuracy hides. I model this as a time-varying availability problem grounded in heterogeneous nucleation theory, evaluated across BDD100K, KITTI, and nuScenes, producing a design tool that links environmental conditions to required thermal mitigation.