MS Mechatronics, Robotics, and Automation Engineering at NYU Tandon (May 2026). First author on an arXiv preprint under peer review at IEEE CASE 2026 and IEEE Robotics and Automation Letters. Production C++17 perception pipeline shipped end-to-end with Docker, docs site, and 162 + 31 tests. Six compute platforms in production code, from 8-core Parallax Propeller cogs written from datasheets to NVIDIA GPU clusters I co-architected. Based in New York City. F-1 / STEM-OPT eligible.

terra-perceive Production C++17 perception pipeline for off-road / construction-site autonomy. From-scratch sector RANSAC, LiDAR-inertial SLAM with Lie groups + on-manifold IMU pre-integration, factor-graph LM, Scan Context, SORT + IMM cascade, probabilistic traversability, 1D CBF safety supervisor. 0.577 m ATE on RELLIS-3D Seq 00 (within 13% of Cartographer). 162 C++ + 31 Python tests, Docker repro in 45 s. M1–M13 shipped. Blog · solo.	ml_inference · arXiv 2603.28890 Bootstrap depth fusion under hardware sensor failure: the surviving valid ToF pixels calibrate a learned monocular depth student to metric scale at runtime. Distilled student at 218 FPS / 4.6 ms / 2.7 GB on Jetson Orin Nano. Costmap obstacle coverage +55% (corridor), +110% (LILocBench) over LiDAR alone. 9/10 closed-loop navigation success matching ground-truth-depth baseline. Under peer review at IEEE CASE 2026 / RA-L.	NCHSB ROS 2 Humble high-speed Ackermann racing platform — modified Traxxas Maxx 4S, custom URDF resolving the four-bar-linkage closed-loop problem, custom C++ Nav2 GlobalPlanner plugin, CVXPY QP racing-line optimizer, Nav2 MPPI tuned for high-speed Ackermann, SLAM Toolbox + EKF retuned away from differential-drive defaults. Emergency-MPC fallback with Human > Object > Wall ethical cost hierarchy. 3-person team (first author).
Unitree Go2 RL — Isaac Lab PPO Walking policy trained in NVIDIA Isaac Lab. Linear velocity tracking 88.0 / angular 47.2 against course targets of 48 / 24 — nearly 2×. Raibert heuristic gait clock, bonus actuator-friction randomization for sim-to-real robustness, RayCaster height-scanner terrain extension. Trained at 4096 parallel environments, 500 epochs in ~12 min on the same NYU AWS ParallelCluster I co-architected. Physical Go2 walked one minute without falling on first deployment. 3-person team.	ChoiRbot swarm 3-robot heterogeneous swarm: two physical bots on radically different MCUs — Arduino Nano 33 BLE Sense + 8-core Parallax Propeller P8X32A bridged through an ESP32 BLE-UART — coordinated through a Raspberry Pi 5 central station running ROS 2 Jazzy. Solo-built firmware on all four platforms; the Propeller multi-cog C was written from datasheets directly. 0.31% EKF position error. Sim-to-real story: Lyapunov controller failed under real motor friction, retuned to PID. Co-implemented Distributed Simplex task assignment and CBF safety filter.	Quantum Geometric Intelligence Lab Two semesters in NYU Finance and Risk Engineering on commodity forecasting. Sem 1: co-authored survey, solo-ablated CNN / CNN+attention / RCNN+GRU / RCNN+attention on multivariate wheat futures. Sem 2: pivoted from regression to classification reformulation (buy/sell/hold signals) and fused FRED-MD macro + FinBERT NLP sentiment + Google Earth Engine NDVI satellite into a unified input space feeding RCNN-with-skip, RCNN-with-attention, BiGRU.

I write production code on every layer of the autonomy stack — six distinct compute platforms in shipped projects, from datasheet-level firmware to GPU-cluster training.

Pushparaju, N., Mattam, V., Arab, A. (2026). Bootstrap Perception Under Hardware Depth Failure for Indoor Robot Navigation. arXiv preprint 2603.28890 — under peer review at IEEE CASE 2026 and IEEE Robotics and Automation Letters.

A self-referential failure-aware sensing hierarchy: the surviving valid ToF pixels (up to 78% invalid in the worst-case reflective-corridor segment) calibrate a learned monocular depth estimator (DA3-Small student, distilled from DA3-Large) to metric scale at runtime — the system fills its own gaps without external data. LiDAR remains the geometric anchor; hardware ToF depth is preserved where valid; learned depth enters only where ToF has dropped out; semantics refine inflation only when needed. Counter-intuitive bonus result: learned depth reduces costmap flicker (0.61% → 0.43% normalized occupancy jitter) by stabilizing intermittent ToF dropouts.

arXiv · Hugging Face models (vortex-depth-v5-general, vortex-depth-v6-pretrained, vortex-depth-v9-corridor) · Code under double-blind review embargo until publication.

Pushparaju, N. et al. (2025). Survey of Commodity Price Forecasting. NYU Quantum Geometric Intelligence Lab — 5-author internal lab deliverable. My sections: data sourcing (futures + USDA WASDE + horticulture panels), feature engineering (continuous-futures construction, stationarity transforms, lag selection via differential evolution and ACF/PACF), CNN architectures for time series, and attention-based RNN/CNN variants with gradient-flow analysis. PDF in repo.

3 merged pull requests to OpenVINO — Intel's production AI inference toolkit — across C++, Node.js, and Python language bindings. Notable: import_model(tensor) JavaScript API addition; segment_mean_csr converter implementation. ~10 PRs in flight across language bindings.

Other public activity: GPU-accelerated NYC taxi data analysis (1.1B records via NVIDIA RAPIDS / cuGraph); modified DBA-Fusion to support stereo cameras; fork of openxrlab/xrdslam for deep-learning SLAM exploration.

Robotics & SLAM — ROS 2 (Humble / Jazzy / Foxy), Nav2 (MPPI, Smac Hybrid A*, behavior trees, lifecycle nodes), SLAM Toolbox, robot_localization EKF, ros2_control, ackermann_steering_controller, URDF / Xacro, Gazebo Harmonic / Fortress, NVIDIA Isaac Lab, KISS-ICP, IMU pre-integration on SO(3), Scan Context loop closure, g2o factor graphs, SuperPoint + SuperGlue, CosPlace, ORB-SLAM3, Hybrid A* / Reeds-Shepp planning.

Machine learning & edge ML — PyTorch, TensorFlow, knowledge distillation (EfficientViT-B1 student from DA3-Large), PPO reinforcement learning, time-series ML with leakage-free TimeSeriesSplit CV, multimodal feature fusion (FRED-MD macro + FinBERT NLP + GEE NDVI satellite), TensorRT FP16 / INT8, ONNX, OpenVINO (3 merged PRs), Apptainer / Singularity for HPC inference.

Embedded firmware (six platforms) — Parallax Propeller P8X32A multi-cog C written from datasheets · Arduino Nano 33 BLE Sense (BLE GATT, encoders, PWM, H-bridge) · ESP32-WROOM-32 (BLE-UART bridges) · Raspberry Pi 5 (Ubuntu 24.04 + ROS 2 Jazzy) · Jetson Orin Nano (TensorRT inference) · NVIDIA GPU clusters (CUDA-aware Slurm jobs).

Cloud & HPC — AWS ParallelCluster co-architected with NYU IT Research Technology Services, CloudFormation IaC, Slurm (interactive + batch), FSx for Lustre, Apptainer SIF containers (deliberate choice over Docker for HPC cgroup integration), Cloudflare Tunnels for zero-trust networking, IP-restricted security groups + IAM hardening. Built the cluster, then used it to train the Go2 PPO policy at 4096 envs and serve an SDXL Turbo image-generation pipeline within a $2,000 AWS budget.

Math & optimization — Eigen3, CVXPY, qpSWIFT, OSQP, Distributed Simplex, Control Barrier Functions, EKF, factor-graph optimization (g2o; GTSAM exposure), Lie groups SO(3) / SE(3).

Languages — C++17, Python 3, C, MATLAB, R, JavaScript (Node.js), SQL, Bash, LaTeX, KUKA KRL.

Industrial robotics — KUKA KR-6, KUKA KRL programming, mastering / tool calibration / base calibration, Schmalz high-flow vacuum end-effectors, offline programming in KUKA Sim Pro and RoboDK.

Data engineering (Cognizant / Sanofi, 1 year 8 months) — Informatica PowerCenter, Oracle 11g/12c, MySQL, Cognos BI, Power BI — multi-region (APAC / EMEA / AMER) ETL with reports informing quarterly business decisions for Sanofi senior leadership under GxP-adjacent pharmaceutical data governance.

May 2026: finishing the MS, shipping nuScenes cross-domain validation (M14) for terra-perceive, IEEE CASE peer-review cycle for the arXiv preprint, wrapping the Sem 2 commodity-classification report at the VIP lab, queuing trt-bench (C++ TensorRT benchmarking utility) as the next public artifact. Looking for robotics engineering · perception engineering · edge-ML deployment · robotics systems integration roles in the United States. Strongest fit in AV deployment teams, defense robotics, warehouse / logistics autonomy, drone autonomy, and edge-ML inference infrastructure. EAD active July 2026; STEM-OPT eligible through ~2029.

New York University, Tandon School of Engineering — M.S. Mechatronics, Robotics, and Automation Engineering (Sep 2024 – May 2026). MS project advised by Prof. Aliasghar Arab (NYU MAE). Coursework: Robot Perception · Reinforcement Learning for Robotics · Networked Robotics Systems / Cooperative Control / Swarming · Advanced Mechatronics · Linear Algebra · Foundations of Robotics · Optimization Techniques · Data-Driven Decision Making · SQP Optimization for Robotics.

SRM Institute of Science and Technology — B.Tech Mechatronics, Robotics, and Automation Engineering (Aug 2017 – Jun 2021). Capstone "Automatic Handling of Cut Textiles" in collaboration with Institut für Textiltechnik (ITA), RWTH Aachen University — pick-and-place automation of limp denim using a KUKA KR-6 with Schmalz high-flow vacuum gripper.

• Email: nishantpushparaju@gmail.com · np3129@nyu.edu (active until May 2026)
• LinkedIn: linkedin.com/in/nishant-pushparaju-857695188
• Portfolio: nishantpushparaju.dev
• Based in New York City · open to relocation anywhere in the United States