I work where perception meets control — where a model has to look at the world and act on it before the next frame lands.
My focus is computer vision, mobile robotics, perception, and control. The things I'm most proud of: a multi-object tracker that follows two people walking around a room without losing track of who's who, RTAB-Map running onboard a snake-inspired robot, and a deep-RL racing agent that beat its classical baseline by 2× on F1TENTH.
The rest of my training fills in around that core — reinforcement learning, classical control, manipulator robotics, biomechanics, and human-robot interaction. (Yes, I built a Cozmo robot that plays Rock-Paper-Scissors and trash-talks the loser.)
Before grad school I was a research intern at DRDL Hyderabad working on actuator control. Outside the lab: Formula 1, Marvel marathons, exploring Boston.
Systematic empirical comparison of PPO, SAC, and TD3 against classical baselines (Pure Pursuit, reactive LiDAR) in the F1TENTH Gym simulator. Five evaluation axes: sample efficiency, peak performance, multi-seed stability, generalization to unseen tracks, and reward-shaping ablations.
Implementation of a Gaussian Process-guided MPPI controller. A Sparse GP occupancy model built from LiDAR estimates both occupancy and uncertainty — high-variance regions become subgoals that guide MPPI beyond its planning horizon, escaping local-minima traps that hinder vanilla MPPI.
Deployed RTAB-Map on COBRA, an 11-joint sidewinding robot. Stack: Intel RealSense D435i feeding a Jetson Orin NX. Benchmarked odometry against OptiTrack ground truth using Evo. Documented graceful degradation during fast sidewinding gaits.
Tracking-by-detection pipeline built from scratch. YOLOv8n provides per-frame detections; a custom Kalman Filter predicts motion; the Hungarian algorithm performs assignment. Supports both IoU and Euclidean cost matrices.
Within-subjects user study (N=24) evaluating how robot personality modulates user perception during a synchronous Rock-Paper-Scissors game with an Anki Cozmo. Custom Python pipeline using MediaPipe Hands for real-time gesture recognition. Three personality conditions.
End-to-end pipeline for muscle-driven simulation of stair ambulation using OpenSim 4.5 and MATLAB. Processed kinematic and kinetic data from the Darmstadt Stair Ambulation Dataset; ran MocoInverse trajectory optimization on the gait2392 model. Validated against EMG.
Modeled a single-axis maglev system from Newton's laws and electromagnetic equations into state-space form. Implemented PID and full-state-feedback controllers in MATLAB / Simulink. Stability analyzed via Lyapunov functions and controllability matrices.
Graduating May 2026. Open to full-time roles across the robotics and applied-ML stack — and equally interested in collaborations, contract work, and research conversations.
Email is fastest — usually within 24 hours.
polagoni.m@northeastern.edu →