Selected work

Representative Papers

Safe On-Orbit Dislodging of Deployable Structures via Robust Adaptive MPC

Mar 21, 2025

This work addresses the unstructured, highly uncertain, time-varying, and safety-critical problem of unsticking jammed solar arrays aboard space stations using robotic manipulators. We propose a novel synergistic framework integrating online ensemble member identification with robust adaptive model predictive control (MPC). A hybrid hinge-based high-fidelity dynamic model enables joint optimization of parameter-space exploration and closed-loop control performance. The approach achieves successful unsticking in both zero-gravity hardware-in-the-loop and ground experiments, reducing parameter estimation error by 42% and improving control success rate by 31% over state-of-the-art methods—while strictly satisfying hard safety constraints throughout. To our knowledge, this is the first work to embed ensemble-based system identification within a robust adaptive MPC loop, establishing a verifiably safe control paradigm for time-varying space robotic systems.

0 citationsRead paper

Making the unmodulated pyramid wavefront sensor smart. II. First on-sky demonstration of extreme adaptive optics with deep learning

Mar 14, 2025Astronomy & Astrophysics

Unmodulated pyramid wavefront sensors (PWFS) suffer from narrow linear dynamic range, insensitivity to petal-piston modes, and difficulty achieving high-speed operation—fundamental limitations hindering their deployment in extreme adaptive optics (XAO). Method: This work presents the first engineering integration of an unmodulated PWFS with deep learning–based wavefront reconstruction in real astronomical observations. We propose a convolutional neural network (CNN) for nonlinear wavefront estimation, optimized for real-time inference via TensorRT, deployed on the MagAO-X hardware platform, and calibrated using an internal light source with domain-adapted transfer calibration. Contribution/Results: Our approach breaks the classical sensitivity–speed trade-off inherent in modulated PWFS. Experiments demonstrate Strehl ratios matching those of optimized modulated PWFS on bright stars; significantly improved performance on faint stars under strong wind conditions; and stable closed-loop bandwidth exceeding 2 kHz. These results validate the feasibility and superiority of unmodulated PWFS for XAO systems.

0 citationsRead paper

AdaptiveCoPilot: Design and Testing of a NeuroAdaptive LLM Cockpit Guidance System in both Novice and Expert Pilots

Jan 07, 2025

Cognitive overload from multitasking in modern cockpits jeopardizes flight safety. To address this, we propose the first neuroadaptive large language model (LLM)-guided system tailored for aviation: it fuses real-time functional near-infrared spectroscopy (fNIRS) neural signals with behavioral data to construct a dual-driven cognitive load model—simultaneously quantifying working memory and perceptual load. We introduce the first deep integration of fNIRS-based neural feedback with context-aware LLMs, enabling dynamic, load-informed modality switching among visual, auditory, and textual prompts. In a VR-based flight simulation experiment with licensed pilots, the system significantly increased the proportion of time spent within the optimal cognitive load zone (p < 0.01), reduced task completion time, and achieved statistically significant improvements in both working memory and perceptual load regulation (p < 0.01). These results empirically validate the efficacy of our neuro-semantic co-adaptive paradigm for adaptive human–autonomy interaction in safety-critical domains.

0 citationsRead paper

Recent publications

Latest Papers

Safe On-Orbit Dislodging of Deployable Structures via Robust Adaptive MPC

Mar 21, 2025

This work addresses the unstructured, highly uncertain, time-varying, and safety-critical problem of unsticking jammed solar arrays aboard space stations using robotic manipulators. We propose a novel synergistic framework integrating online ensemble member identification with robust adaptive model predictive control (MPC). A hybrid hinge-based high-fidelity dynamic model enables joint optimization of parameter-space exploration and closed-loop control performance. The approach achieves successful unsticking in both zero-gravity hardware-in-the-loop and ground experiments, reducing parameter estimation error by 42% and improving control success rate by 31% over state-of-the-art methods—while strictly satisfying hard safety constraints throughout. To our knowledge, this is the first work to embed ensemble-based system identification within a robust adaptive MPC loop, establishing a verifiably safe control paradigm for time-varying space robotic systems.

0 citationsRead paper

Making the unmodulated pyramid wavefront sensor smart. II. First on-sky demonstration of extreme adaptive optics with deep learning

Mar 14, 2025Astronomy &amp; Astrophysics

Unmodulated pyramid wavefront sensors (PWFS) suffer from narrow linear dynamic range, insensitivity to petal-piston modes, and difficulty achieving high-speed operation—fundamental limitations hindering their deployment in extreme adaptive optics (XAO). Method: This work presents the first engineering integration of an unmodulated PWFS with deep learning–based wavefront reconstruction in real astronomical observations. We propose a convolutional neural network (CNN) for nonlinear wavefront estimation, optimized for real-time inference via TensorRT, deployed on the MagAO-X hardware platform, and calibrated using an internal light source with domain-adapted transfer calibration. Contribution/Results: Our approach breaks the classical sensitivity–speed trade-off inherent in modulated PWFS. Experiments demonstrate Strehl ratios matching those of optimized modulated PWFS on bright stars; significantly improved performance on faint stars under strong wind conditions; and stable closed-loop bandwidth exceeding 2 kHz. These results validate the feasibility and superiority of unmodulated PWFS for XAO systems.

0 citationsRead paper

AdaptiveCoPilot: Design and Testing of a NeuroAdaptive LLM Cockpit Guidance System in both Novice and Expert Pilots

Jan 07, 2025

Cognitive overload from multitasking in modern cockpits jeopardizes flight safety. To address this, we propose the first neuroadaptive large language model (LLM)-guided system tailored for aviation: it fuses real-time functional near-infrared spectroscopy (fNIRS) neural signals with behavioral data to construct a dual-driven cognitive load model—simultaneously quantifying working memory and perceptual load. We introduce the first deep integration of fNIRS-based neural feedback with context-aware LLMs, enabling dynamic, load-informed modality switching among visual, auditory, and textual prompts. In a VR-based flight simulation experiment with licensed pilots, the system significantly increased the proportion of time spent within the optimal cognitive load zone (p < 0.01), reduced task completion time, and achieved statistically significant improvements in both working memory and perceptual load regulation (p < 0.01). These results empirically validate the efficacy of our neuro-semantic co-adaptive paradigm for adaptive human–autonomy interaction in safety-critical domains.

0 citationsRead paper