Nishant Pushparaju

A from-scratch C + +17 perception pipeline for off-road autonomy on the RELLIS-3D dataset. I implemented sector-based RANSAC ground segmentation that handles graded terrain where global RANSAC fails, a PCA-based traversability grid with explicit risk and confidence separation, camera-LiDAR projection with SegFormer semantic fusion, and a from-scratch LiDAR-inertial SLAM backend with SO(3) Lie groups, IMU preintegration, and Scan Context loop closure. Benchmarked against Google Cartographer.

A self-referential perception architecture for indoor navigation. Time-of-flight cameras lose up to 78 % of depth pixels on reflective surfaces; in this work, the surviving valid pixels calibrate a learned monocular depth estimator to metric scale at runtime, so the sensor fills its own gaps without external data. LiDAR remains the anchor, hardware depth is preserved where valid, learned depth enters only where needed. Knowledge distillation from DA3-Large teacher to a 5.31 M-parameter EfficientViT-B1 student. Under peer review at IEEE CASE 2026 and IEEE Robotics & Automation Letters.

The hardware platform behind the data behind the paper. I built a modified Traxxas Maxx 4S 1/10-scale Ackermann RC car for narrow indoor corridor navigation, designed for 30 m/s and validated in Gazebo Fortress to 25 m/s. Custom Nav2 GlobalPlanner C + + plugin. F1-style racing-line optimizer in CVXPY that maximizes corner-exit speed via outside-inside-outside apex geometry. SLAM Toolbox and MPPI controller tuned specifically for Ackermann kinematics, not the differential-drive defaults that ship with Nav2. Camera-LiDAR sensor fusion via HSV segmentation, Farneback optical flow, ego-motion compensation, and 15-frame temporal voting.

I co-architected NYU’s AWS cloud-burst HPC cluster, then deployed a personal SDXL Turbo image-generation pipeline on top of it — and used the same cluster to train the Go2 walking policy. The build-and-use arc is the differentiator. AWS ParallelCluster with CloudFormation IaC, persistent head node, ephemeral GPU compute nodes auto-scaling 0→20 instances on Slurm queue demand, 10-minute idle timeout for cost control. FSx for Lustre at /work for shared project data. Apptainer/Singularity rather than Docker because HPC compute nodes don’t run Docker daemons. SDXL pipeline: personal website trigger → Cloudflare Tunnel → AWS GPU node → SDXL Turbo inference → image returned. Mentored by Shenglong Wang (NYU researcher and HPC engineer) and Stratos Efstathiadis (Senior Director, NYU IT RTS).

I wrote the Parallax Propeller multi-cog C firmware from datasheets — existing community resources did not cover the MPU-6050 + quadrature encoder + ESP32 UART integration we needed on Propeller, so the firmware went straight from datasheet to working code with no intermediate tutorial. Beyond that, I owned the layered software architecture across the heterogeneous platform: ArUco vision-based localization (DICT_7X7_250, solvePnP, Rodrigues-to-quaternion), the BLE bridge from ROS 2 Twist to GATT writes via bleak/asyncio, and the Pandebono Arduino firmware. Sim-to-real engineering insight from this project: a clean Lyapunov controller that worked in Gazebo failed in real-world deployment due to motor friction and unmodeled dynamics — documented the failure mode and retuned to a PID for the velocity command. Textbook sim-to-real gap, observed and resolved.

I factored a hard metric navigation problem into a topological one. CosPlace ResNet-18 descriptors give fast top-K recall over the database; SuperPoint + SuperGlue verifies each candidate with essential-matrix inlier count. Both used for what each is good at. Deployed on the Frodobots remote-rover platform. My project partner refactored the joint codebase into the production basis for his EarthRover Challenge competition entry.

I trained a Unitree Go2 walking policy at 4096 parallel environments on the NYU AWS HPC ParallelCluster I co-architected upstream — using the infrastructure I helped build to train downstream research. Reward shaping with action-rate penalties on first and second derivatives, an explicit PD controller replacing the implicit actuator, a Raibert heuristic gait clock, and a stiction-plus-viscous actuator friction model randomized per episode. Honestly documented one reward term that silently zeroed; the policy still learned a stable trot.

Two semesters in Prof. Aboussalah’s lab. I co-authored a 5-author survey on deep learning for commodity price forecasting; my sections covered data sourcing, feature engineering (continuous futures, stationarity, lag selection via differential evolution), and gradient-flow analysis showing how attention mitigates vanishing gradients. In semester two I pivoted from regression to buy/sell/hold classification and expanded the input space from FRED-MD macros to incorporate FinBERT NLP sentiment and Google Earth Engine NDVI signals over wheat-growing regions.