- MAS Affiliated Graduate Students
Stephen P. Kaputsos is an embodied AI researcher, architect, and engineer specializing in perceptual AI, autonomous systems, sensing architecture, multi-modal sensor fusion, autonomous navigation, and AI/ML. Two Master of Science graduate degrees: MIT in artificial intelligence, robotics and simulation (5.0/5.0 GPA); Johns Hopkins in Systems Engineering (4.0/4.0 GPA). His work covers the full embodied and physical AI stack - data systems, AI perception and cognition development, simulated and physical robotic embodiment - from requirements through sensor selection, model development, deployment, and field validation.
His MIT graduate program research centered on machine perception and autonomous navigation for robotic platforms (AUAVs), spanning the full embodied AI stack. He authored a custom SDK integration extending commercial robotic systems beyond their native control interfaces to enable deep learning offboard perception, then specified and integrated custom multi-modal sensor payloads within the platform size, weight, and power envelope. Perception ran offboard on mobile and server compute against sensor streams from the platform, with an onboard path for comparison. He developed and benchmarked classical and machine learning approaches to autonomous navigation against the same task: a deterministic pipeline coupling marker detection, relative state estimation, and trajectory generation, reliable without training cost but limited as environments grow complex and dynamic; and a deep reinforcement learning (DRL) policy reaching 98.3% task success. He also built SLAM and V-SLAM navigation for state estimation in unmapped environments. He built the autonomous robotics simulation and training platform, integrating the full OpenAI Baselines algorithm library with network hyperparameters configurable through the interface and trained models saved and reloaded automatically, alongside the domain randomization environment generating its synthetic datasets. A recurring thread is the boundary between simulated and physical systems - calibrating physical cameras and matching simulated sensor models and scene geometry to measured values, establishing conformance so that simulation results support real sensing and model decisions.
Beyond MIT, he has shipped perception systems to production under inherited compute and memory budgets, run sensor placement and configuration studies for multi-camera systems operating under heavy occlusion, and built the large-scale data collection, annotation, and schema pipelines feeding model training. He also worked extensively with IEEE and ETSI specifications covering antenna bands, frequency allocations, and RF network standards, giving him substantial depth in electromagnetism and the physics governing sensing modalities across the spectrum. Earlier work includes multi-robot architectures in which experience gathered by any agent accelerates learning across the fleet, spatial computing systems where one perception stack serves both simulated and physical deployments, and systems that adapt autonomous behavior in real time to sensed operator state.
Stephen has led engineering and cross-disciplinary research teams through development and deployment, is a named inventor on a published U.S. patent application filed by Verizon for large language model based device and network management and automation (U.S. Patent Application 2025/0307542 A1), and is a published IEEE conference author. He designed and delivered an embodied AI executive education program at the MIT Media Lab for senior United States Air Force leadership, engineering leads, and stakeholders. He is a graduate of the American Management Association's Strategic Leadership Program and a keynote speaker at MIT Reality Hack, one of the world's largest XR (AR/VR and spatial computing) events.