Selected work

Representative Papers

From Task Allocation to Risk Clearing: A Unifying Interface for Mixed Human-Agent Societies

May 26, 2026

This work addresses the challenge of coordinating heterogeneous agents in safety-critical human–AI collaborative settings, where existing mechanisms struggle to simultaneously support dynamic task allocation, commitment under uncertainty, and scalable integration. The paper introduces Risk-aware Option Clearing (ROC), a novel coordination framework that treats risk-aware options as fundamental units of interaction. Each option encapsulates an agent’s temporally extended skill along with a concise risk summary. A central clearinghouse optimizes task assignment by jointly considering risk-adjusted utility, temporal deadlines, and safety constraints. The framework unifies diverse deployment paradigms—from data-driven learning to full distributional prediction—providing a transparent, interpretable, and scalable coordination infrastructure for mixed human–machine systems and advancing the applicability of risk-aware clearing layers in hybrid social environments.

0 citationsRead paper

AI Disclosure with DAISY

Apr 03, 2026

The increasing use of artificial intelligence (AI) in scientific research is hindered by inconsistent and often inadequate disclosure practices, as authors face social, cognitive, and emotional barriers, and current policies offer insufficient support. This study proposes DAISY, a form-based structured disclosure tool developed through a co-design approach that conceptualizes AI disclosure as a sociotechnical practice. Integrating insights from a literature review, co-design workshops (N=11), and user studies (N=31), DAISY draws on principles from human-computer interaction and responsible AI. Findings demonstrate that DAISY significantly enhances the completeness of disclosure statements and the clarity of AI usage details, while preserving authors’ comfort in reporting. The tool thus offers a scalable, practical solution to improve transparency in research involving AI.

0 citationsRead paper

Neural-Symbolic Integration with Evolvable Policies

Jan 08, 2026arXiv.org

This work proposes a novel evolutionary approach to neural-symbolic integration that overcomes the limitations of existing systems, which typically rely on predefined or differentiable symbolic policies and thus struggle in settings lacking expert knowledge or when policies are non-differentiable. The method treats the neural-symbolic system as an evolvable individual, simultaneously learning non-differentiable symbolic policies and neural network weights from an initial state of an empty policy and random weights. By integrating the NEUROLOG architecture, Valiant’s evolvability framework, Machine Coaching semantics, and abduction-driven neural training, the system evolves through mutation and fitness-based selection to approximate the target policy. Experimental results demonstrate that the approach can effectively learn complex symbolic strategies without any prior knowledge, achieving near-perfect median accuracy (~100%) and significantly extending the applicability of neural-symbolic systems to scenarios without expert guidance.

0 citationsRead paper

PEDESTRIAN: An Egocentric Vision Dataset for Obstacle Detection on Pavements

Dec 22, 2025

Urban sidewalks are frequently obstructed by hazards that compromise pedestrian safety, yet real-time detection is hindered by the absence of high-quality, multi-class egocentric visual datasets. To address this gap, we introduce the first large-scale egocentric video dataset specifically designed for sidewalk obstacle detection—comprising 340 real-world smartphone-recorded videos spanning 29 common obstacle categories. We systematically define and publicly release a high-fidelity, fine-grained annotation benchmark, the first of its kind, thereby filling a critical void in open pedestrian safety resources. Leveraging this dataset, we conduct a comprehensive evaluation of state-of-the-art object detectors—including YOLOv8 and Mask R-CNN—establishing fully reproducible baselines. Our best-performing model achieves a mean average precision (mAP@0.5) of 68.3%. This work provides both an essential data foundation and an authoritative performance benchmark for developing robust pedestrian safety warning systems.

0 citationsRead paper

GAEA: Experiences and Lessons Learned from a Country-Scale Environmental Digital Twin

Nov 17, 2025

This study synthesizes three years of operational experience with GAEA, Cyprus’s national environmental digital twin system. Addressing cross-sectoral environmental data silos and insufficient dynamic decision support, the project developed a high-fidelity, island-wide digital twin platform integrating 27 geospatial services. It fuses heterogeneous, multi-source environmental data and enables near-real-time updates via geospatial analytics, environmental modeling, and a cloud-native architecture. Its key contribution lies in achieving the first nationally scaled, long-term operational deployment of an environmental digital twin—demonstrating a replicable paradigm for cross-domain data integration, service interoperability, and collaborative governance. The system has already supported evidence-based decision-making in urban planning, agricultural management, and insurance actuarial modeling, markedly improving environmental data utilization efficiency and policy responsiveness. It provides a transferable technical framework and institutional blueprint for smart environmental governance in small- and medium-sized nations.

0 citationsRead paper

Recent publications

Latest Papers

From Task Allocation to Risk Clearing: A Unifying Interface for Mixed Human-Agent Societies

May 26, 2026

This work addresses the challenge of coordinating heterogeneous agents in safety-critical human–AI collaborative settings, where existing mechanisms struggle to simultaneously support dynamic task allocation, commitment under uncertainty, and scalable integration. The paper introduces Risk-aware Option Clearing (ROC), a novel coordination framework that treats risk-aware options as fundamental units of interaction. Each option encapsulates an agent’s temporally extended skill along with a concise risk summary. A central clearinghouse optimizes task assignment by jointly considering risk-adjusted utility, temporal deadlines, and safety constraints. The framework unifies diverse deployment paradigms—from data-driven learning to full distributional prediction—providing a transparent, interpretable, and scalable coordination infrastructure for mixed human–machine systems and advancing the applicability of risk-aware clearing layers in hybrid social environments.

0 citationsRead paper

AI Disclosure with DAISY

Apr 03, 2026

The increasing use of artificial intelligence (AI) in scientific research is hindered by inconsistent and often inadequate disclosure practices, as authors face social, cognitive, and emotional barriers, and current policies offer insufficient support. This study proposes DAISY, a form-based structured disclosure tool developed through a co-design approach that conceptualizes AI disclosure as a sociotechnical practice. Integrating insights from a literature review, co-design workshops (N=11), and user studies (N=31), DAISY draws on principles from human-computer interaction and responsible AI. Findings demonstrate that DAISY significantly enhances the completeness of disclosure statements and the clarity of AI usage details, while preserving authors’ comfort in reporting. The tool thus offers a scalable, practical solution to improve transparency in research involving AI.

0 citationsRead paper

Neural-Symbolic Integration with Evolvable Policies

Jan 08, 2026arXiv.org

This work proposes a novel evolutionary approach to neural-symbolic integration that overcomes the limitations of existing systems, which typically rely on predefined or differentiable symbolic policies and thus struggle in settings lacking expert knowledge or when policies are non-differentiable. The method treats the neural-symbolic system as an evolvable individual, simultaneously learning non-differentiable symbolic policies and neural network weights from an initial state of an empty policy and random weights. By integrating the NEUROLOG architecture, Valiant’s evolvability framework, Machine Coaching semantics, and abduction-driven neural training, the system evolves through mutation and fitness-based selection to approximate the target policy. Experimental results demonstrate that the approach can effectively learn complex symbolic strategies without any prior knowledge, achieving near-perfect median accuracy (~100%) and significantly extending the applicability of neural-symbolic systems to scenarios without expert guidance.

0 citationsRead paper

PEDESTRIAN: An Egocentric Vision Dataset for Obstacle Detection on Pavements

Dec 22, 2025

Urban sidewalks are frequently obstructed by hazards that compromise pedestrian safety, yet real-time detection is hindered by the absence of high-quality, multi-class egocentric visual datasets. To address this gap, we introduce the first large-scale egocentric video dataset specifically designed for sidewalk obstacle detection—comprising 340 real-world smartphone-recorded videos spanning 29 common obstacle categories. We systematically define and publicly release a high-fidelity, fine-grained annotation benchmark, the first of its kind, thereby filling a critical void in open pedestrian safety resources. Leveraging this dataset, we conduct a comprehensive evaluation of state-of-the-art object detectors—including YOLOv8 and Mask R-CNN—establishing fully reproducible baselines. Our best-performing model achieves a mean average precision (mAP@0.5) of 68.3%. This work provides both an essential data foundation and an authoritative performance benchmark for developing robust pedestrian safety warning systems.

0 citationsRead paper

GAEA: Experiences and Lessons Learned from a Country-Scale Environmental Digital Twin

Nov 17, 2025

This study synthesizes three years of operational experience with GAEA, Cyprus’s national environmental digital twin system. Addressing cross-sectoral environmental data silos and insufficient dynamic decision support, the project developed a high-fidelity, island-wide digital twin platform integrating 27 geospatial services. It fuses heterogeneous, multi-source environmental data and enables near-real-time updates via geospatial analytics, environmental modeling, and a cloud-native architecture. Its key contribution lies in achieving the first nationally scaled, long-term operational deployment of an environmental digital twin—demonstrating a replicable paradigm for cross-domain data integration, service interoperability, and collaborative governance. The system has already supported evidence-based decision-making in urban planning, agricultural management, and insurance actuarial modeling, markedly improving environmental data utilization efficiency and policy responsiveness. It provides a transferable technical framework and institutional blueprint for smart environmental governance in small- and medium-sized nations.

0 citationsRead paper