Selected work

Representative Papers

Randomized Atomic Feature Models for Physics-Informed Identification of Dynamic Systems

May 14, 2026

This work addresses the challenge of identifying dynamic systems under insufficient excitation, where achieving physical consistency, stability, and interpretability simultaneously remains difficult. To this end, we propose a physics-informed framework based on stochastic stable atomic features. The impulse response is modeled as a random superposition of damped complex exponential atoms, and physically interpretable modal parameters are efficiently recovered through convex regularized least squares subject to explicit stability constraints. By integrating perspectives from Disk–Bochner operators, reproducing kernel Hilbert space (RKHS) theory, and the Kalman–Yakubovich–Popov (KYP) lemma, the method embeds engineering priors—such as stability and DC gain—into a finite-dimensional optimization framework. Experimental results demonstrate that the proposed approach significantly improves identification accuracy under poorly excited conditions while rigorously preserving system stability and structural interpretability.

0 citationsRead paper

Path Integral Particle Filtering for Hybrid Systems via Saltation Matrices

Mar 01, 2026

This work addresses the challenge of state estimation in hybrid systems subject to intermittent contact, where non-Gaussian noise and abrupt dynamic shifts degrade performance. To this end, we propose a novel path integral filtering method that integrates a jump matrix to explicitly model the discontinuous dynamics and uncertainty propagation induced by contact events. By reformulating the state smoothing problem as an optimal control problem, our approach embeds the jump matrix into the path integral framework for the first time, yielding a new particle filtering algorithm tailored for stochastic hybrid systems. This integration significantly enhances robustness against outliers and non-Gaussian disturbances. Experimental results demonstrate that the proposed method consistently outperforms strong existing baselines across multiple scenarios, achieving efficient and reliable state estimation.

0 citationsRead paper

TITAN: Twin-Informed Topology Adaptation for LAWN-enabled D2C Communication

Feb 28, 2026

This study addresses the unreliability of satellite-to-ground direct communication in dense urban environments, where signal blockage often compromises post-disaster connectivity. To overcome this challenge, the authors propose leveraging unmanned aerial vehicles (UAVs) within a low-altitude wireless network (LAWN) as adaptive aerial relays for direct-to-cellular (D2C) links. The work innovatively integrates high-fidelity digital twins with Sionna-based ray-tracing channel modeling and employs Bayesian optimization to dynamically configure the relay topology, thereby surpassing the limitations of conventional heuristic or statistical channel approaches. System-level simulations demonstrate significant performance gains over existing methods, including a 32.2% improvement in user coverage, a 64.9% increase in aggregate system throughput, and a 49.3% enhancement in fairness.

0 citationsRead paper

SNAP: Towards Segmenting Anything in Any Point Cloud

Oct 13, 2025

Existing interactive 3D point cloud segmentation methods are constrained by single-scene-domain applicability (e.g., indoor-only or outdoor-only) and single-modality interaction (point prompts only or text prompts only); moreover, multi-dataset joint training often induces negative transfer, severely limiting generalization. This paper proposes the first universal point cloud segmentation framework supporting dual-modal interaction—both point and text prompts. To mitigate negative transfer, we introduce domain-adaptive normalization. Furthermore, we integrate CLIP-based text embedding matching to automatically generate mask proposals, enabling open-vocabulary understanding and panoptic segmentation. The model is jointly trained on seven cross-domain datasets and achieves state-of-the-art performance on eight of nine zero-shot spatial prompting benchmarks, while remaining competitive across all five text-prompting benchmarks. Our approach significantly enhances cross-domain and cross-modal generalization, as well as practical usability.

0 citationsRead paper

Dynamic read & write optimization with TurtleKV

Sep 12, 2025

Traditional key-value stores face a fundamental trade-off among memory overhead, read performance, and write performance—a “trilemma” that limits adaptability. This paper introduces TurtleKV, the first KV system enabling *online, dynamic read-write performance tuning*. Its core innovations are: (1) a bias-free on-disk data structure that eliminates the write bias inherent in LSM-trees; and (2) a fine-grained, runtime memory allocation mechanism that enables bidirectional, real-time optimization of read and write throughput. Evaluated on YCSB, TurtleKV achieves, at comparable space amplification, 8× higher write throughput and 5× higher read throughput than RocksDB; versus SplinterDB, it delivers 40% faster point lookups, 6× faster range scans, and 50% lower space amplification. TurtleKV is the first system to jointly optimize read and write performance under tight space amplification constraints while supporting online, adaptive configuration.

0 citationsRead paper

Recent publications

Latest Papers

Randomized Atomic Feature Models for Physics-Informed Identification of Dynamic Systems

May 14, 2026

This work addresses the challenge of identifying dynamic systems under insufficient excitation, where achieving physical consistency, stability, and interpretability simultaneously remains difficult. To this end, we propose a physics-informed framework based on stochastic stable atomic features. The impulse response is modeled as a random superposition of damped complex exponential atoms, and physically interpretable modal parameters are efficiently recovered through convex regularized least squares subject to explicit stability constraints. By integrating perspectives from Disk–Bochner operators, reproducing kernel Hilbert space (RKHS) theory, and the Kalman–Yakubovich–Popov (KYP) lemma, the method embeds engineering priors—such as stability and DC gain—into a finite-dimensional optimization framework. Experimental results demonstrate that the proposed approach significantly improves identification accuracy under poorly excited conditions while rigorously preserving system stability and structural interpretability.

0 citationsRead paper

Path Integral Particle Filtering for Hybrid Systems via Saltation Matrices

Mar 01, 2026

This work addresses the challenge of state estimation in hybrid systems subject to intermittent contact, where non-Gaussian noise and abrupt dynamic shifts degrade performance. To this end, we propose a novel path integral filtering method that integrates a jump matrix to explicitly model the discontinuous dynamics and uncertainty propagation induced by contact events. By reformulating the state smoothing problem as an optimal control problem, our approach embeds the jump matrix into the path integral framework for the first time, yielding a new particle filtering algorithm tailored for stochastic hybrid systems. This integration significantly enhances robustness against outliers and non-Gaussian disturbances. Experimental results demonstrate that the proposed method consistently outperforms strong existing baselines across multiple scenarios, achieving efficient and reliable state estimation.

0 citationsRead paper

TITAN: Twin-Informed Topology Adaptation for LAWN-enabled D2C Communication

Feb 28, 2026

This study addresses the unreliability of satellite-to-ground direct communication in dense urban environments, where signal blockage often compromises post-disaster connectivity. To overcome this challenge, the authors propose leveraging unmanned aerial vehicles (UAVs) within a low-altitude wireless network (LAWN) as adaptive aerial relays for direct-to-cellular (D2C) links. The work innovatively integrates high-fidelity digital twins with Sionna-based ray-tracing channel modeling and employs Bayesian optimization to dynamically configure the relay topology, thereby surpassing the limitations of conventional heuristic or statistical channel approaches. System-level simulations demonstrate significant performance gains over existing methods, including a 32.2% improvement in user coverage, a 64.9% increase in aggregate system throughput, and a 49.3% enhancement in fairness.

0 citationsRead paper

SNAP: Towards Segmenting Anything in Any Point Cloud

Oct 13, 2025

Existing interactive 3D point cloud segmentation methods are constrained by single-scene-domain applicability (e.g., indoor-only or outdoor-only) and single-modality interaction (point prompts only or text prompts only); moreover, multi-dataset joint training often induces negative transfer, severely limiting generalization. This paper proposes the first universal point cloud segmentation framework supporting dual-modal interaction—both point and text prompts. To mitigate negative transfer, we introduce domain-adaptive normalization. Furthermore, we integrate CLIP-based text embedding matching to automatically generate mask proposals, enabling open-vocabulary understanding and panoptic segmentation. The model is jointly trained on seven cross-domain datasets and achieves state-of-the-art performance on eight of nine zero-shot spatial prompting benchmarks, while remaining competitive across all five text-prompting benchmarks. Our approach significantly enhances cross-domain and cross-modal generalization, as well as practical usability.

0 citationsRead paper

Dynamic read & write optimization with TurtleKV

Sep 12, 2025

Traditional key-value stores face a fundamental trade-off among memory overhead, read performance, and write performance—a “trilemma” that limits adaptability. This paper introduces TurtleKV, the first KV system enabling *online, dynamic read-write performance tuning*. Its core innovations are: (1) a bias-free on-disk data structure that eliminates the write bias inherent in LSM-trees; and (2) a fine-grained, runtime memory allocation mechanism that enables bidirectional, real-time optimization of read and write throughput. Evaluated on YCSB, TurtleKV achieves, at comparable space amplification, 8× higher write throughput and 5× higher read throughput than RocksDB; versus SplinterDB, it delivers 40% faster point lookups, 6× faster range scans, and 50% lower space amplification. TurtleKV is the first system to jointly optimize read and write performance under tight space amplification constraints while supporting online, adaptive configuration.

0 citationsRead paper