AI Autonomy Demands Graduated Oversight, Not Binary Control

Framework operationalizing Berlyne's collative variables (novelty, complexity, conflict, uncertainty) as adaptive linguistic interventions for LLM-based conversational tutoring. Tested across 270 tutoring conversations spanning multiple model families and domains, producing up to 2.4x more conversational turns under fixed time budgets through curiosity-oriented interventions.

Machine-knittable e-textile prototyping platform using industrial digital knitting of conductive yarn to integrate power and communication buses into large-scale textiles. Employs soft-magnetic connectors for repeatable sensor attachment without fabric damage, plus LED-positioning system for automatic sensor node identification and localization. Demonstrated through forearm movement calibration and position-aware temperature mapping.

Human-in-the-loop segmentation framework for pulmonary nodules built on SAM. Humans iteratively refine prompts through trial-and-error learning and semantic reasoning to guide SAM toward higher-quality masks. Validated on chest CT scans from 1,179 patients across 12 centers, achieving mean Dice score of almost 85%, outperforming five state-of-the-art deep learning models by 10-22%.

MIT-licensed swipe corpus containing over 1M donated swipes from more than 12k donor sessions. Released to support training of layout-agnostic neural swipe decoders. Largest MIT-licensed swipe corpus available, designed to enable generalization beyond English QWERTY layouts through geometric augmentations of swipe trajectories and keyboard layouts.

Layout-agnostic neural swipe decoder that predicts character indication and spatial location at each swipe point. Keyboard layout supplied at inference time rather than learned during training. Trained models generalize to layouts absent from training, in some cases more accurately than the training layout. Combines layout-flexibility of algorithmic decoder with neural model accuracy. Publicly available trained models.

AR-HRC system integrating shared semantic environment mapping across AR headset and robot, context-aware world-in-miniature representation for path authoring, and VLM-driven AR view management that adapts virtual content color, size, and orientation for legibility. Implemented with Meta Quest 3 and Unitree Go2 quadruped. User study (N=13) on 30.5m outdoor inspection task showed 66% faster completion times and significantly lower workload.

AR system for warehouse HRC enabling real-time robot situated analytics and control with obstruction detection and mitigation pipeline (ODM) uniquely suited for AR-HRC. Preserves real-world visibility while displaying robot feedback. User study (N=34) showed 46% increase in efficiency on overall HRC task and 61% faster performance on subtasks requiring real-world visibility when ODM was active.

Gamified abstraction of HRC warehouse inventory picking for controlled evaluation. Virtual monsters serve as proxies for pick targets while labeled and object-marked boxes preserve real-world identification demands of picking tasks. Used to evaluate AR-HRC systems in warehouse scenarios with 34 participants.