Arnav Chopra

Test Execution Engine: Distributed Script Execution Platform

Built a Flask-based platform for centrally managing and executing Python test scripts across registered devices. Implemented script upload and validation with paired test files, expected-vs-actual output comparison, timestamped archival, device metadata management, execution output/error capture, and persistent result history.

Extended the initial file-based prototype into a SQLAlchemy/SQLite-backed implementation with a device and script registry, job queue states, and batch execution that expands N selected tests across M selected devices into N×M tracked jobs. The work was demonstrated for a proprietary customer environment; public screenshots and video are not available.

Desktop Toybot: Multi-board Autonomous Robot

Designed a compact 3D-printed chassis in SolidWorks and integrated a Raspberry Pi and NVIDIA Jetson Orin Nano with sensors, actuators, display hardware, computer vision, and voice/AI services. Iterated the enclosure after identifying thermal constraints and validated stable extended operation.

Physical Inverted Pendulum Control

Modeled and linearized an inverted-pendulum system, designed a discrete-time compensator, and implemented the controller on Arduino using an AS5600 encoder and motor drive. Built separate calibration and hardware test routines and added a safety shutdown for excessive angular deviation.

Mobile Robot Localization & Navigation

Built a localization and navigation stack combining wheel odometry with RPLiDAR scan matching and probabilistic state estimation. Integrated embedded sensor/motor I/O with an external Python/C++ compute layer and validated closed-loop navigation across varied indoor floor conditions.

Quadruped Gait Stabilization via Reinforcement Learning

Trained and evaluated a PPO locomotion policy for a Unitree Go2 in NVIDIA Isaac Lab using foot-placement heuristics, torque-level PD control, domain randomization, and distributed training on the NYU Greene HPC cluster.

3D Quadrotor Stabilization & Deployment Pipeline

Developed a Soft Actor-Critic policy for 3D quadrotor position control in MuJoCo, then exported the learned policy to ONNX and integrated it with a ROS2 interface. Validated the same observation/action conventions across MuJoCo, a simplified physics backend, and PX4 software-in-the-loop.

Model Predictive Control for Vehicle Path Tracking

Implemented a linear time-varying MPC controller for a bicycle-model vehicle in MATLAB. Linearized the model along reference trajectories, assembled receding-horizon prediction matrices, solved constrained quadratic programs, and compared tracking behavior with a PID baseline.

Triple Inverted Pendulum via Reinforcement Learning

Modeled a three-link cart-pendulum system and trained a PPO policy for near-upright stabilization. Improved learning by diagnosing reward exploitation, reshaping the objective, and introducing curriculum-based initial-condition randomization rather than relying only on hyperparameter tuning.

Autonomous Flappy Bird with Kalman State Estimation

Built a physics-aware Kalman filter to estimate vertical position and velocity from noisy position measurements, then used the estimated state in a lookahead controller that compared flap versus wait trajectories before each action.

VLM Safety Evaluation: OOPS Framework

Co-developed a benchmark for spatial reasoning and safety compliance in Vision-Language Models, stress-testing multiple models against scene anomalies and performing error analysis to identify failure modes and overconfidence patterns relevant to navigation.

Computer Vision Detection & Evaluation

Developed object-detection and evaluation workflows using YOLOv8 and OpenCV, with emphasis on reproducible testing across lighting, occlusion, and viewpoint changes rather than reporting a single best-case accuracy result.