Selected work

Representative Papers

FastDSAC: Enhancing Policy Plasticity via Constrained Exploration for Scalable Humanoid Locomotion

Jun 30, 2026

This work addresses the instability in value estimation and degradation of policy plasticity commonly observed in high-throughput sampling scenarios with frequent data updates. To mitigate these issues, the authors propose FastDSAC, an algorithm built upon a distributed Actor-Critic framework that models the policy using a truncated Gaussian distribution to simultaneously respect action constraints and preserve exploratory stochasticity. The method incorporates an adaptive variance modulation mechanism to enhance the accuracy of value estimation and employs implicit regularization to maintain the adaptability of the policy network. Experimental results demonstrate that FastDSAC achieves more stable training dynamics, faster convergence, and superior asymptotic performance compared to existing approaches on both the MuJoCo Playground and HumanoidBench benchmarks.

0 citationsRead paper

UAV Trajectory and Bandwidth Allocation for Efficient Data Collection in Low-Altitude Intelligent IoT: A Hierarchical DRL Approach

Apr 25, 2026

This work addresses the challenges of unknown interference, dynamic data volumes, and limited onboard computational resources in unmanned aerial vehicle (UAV)-based data collection within low-altitude intelligent Internet of Things (IoT) systems. To tackle these issues, the authors propose a lightweight hierarchical deep reinforcement learning framework, termed TBH-DDPG, which jointly optimizes data collection efficiency through coarse-grained trajectory planning at the upper level and fine-grained bandwidth allocation at the lower level. Evaluated in realistic environments featuring interference sources, time-varying data demands, and diverse obstacles, the proposed method achieves rapid convergence and low computational overhead. Compared to non-hierarchical approaches, it improves convergence speed by 44.44% and reduces computational cost by 58.05%, significantly enhancing total data throughput and overall system performance.

0 citationsRead paper

Revisiting and Expanding the IPv6 Network Periphery: Global-Scale Measurement and Security Analysis

Apr 21, 2026

This study addresses structural security risks at the IPv6 network edge—such as service exposure and routing loops—for which global-scale systematic assessment has been lacking. The authors present the first comprehensive IPv6 edge security measurement covering 73 countries, introducing a Response-Guided Prefix Selection (RGPS) strategy to efficiently scan high-value targets. They further develop a Hierarchical Large Language Model Exposure Verification (HLEV) framework to analyze unauthorized access risks. Their analysis identifies 281.9 million active IPv6 edge nodes, with a service exposure rate of 2.5%, and detects 4.5 million routing loop responses. The work also uncovers multiple security vulnerabilities in widely used LLM deployment tools, stemming from missing authentication mechanisms due to insecure default configurations.

0 citationsRead paper

Position-Prior-Guided Network for System Matrix Super-Resolution in Magnetic Particle Imaging

Nov 08, 2025

To address the time-consuming and frequently repeated system matrix (SM) calibration in magnetic particle imaging (MPI), this paper proposes a physics-informed deep super-resolution method. The core innovation lies in the first incorporation of SM’s intrinsic spatial symmetry and positional prior into the network architecture, realized via a position-guided convolutional module that jointly enforces physical constraints and data-driven learning. The method enables 2D and 3D SM reconstruction without additional hardware or measurements. Experiments demonstrate that, at identical downsampling ratios, our approach achieves ≥2.1 dB higher PSNR and ≥0.03 higher SSIM than purely data-driven baselines, reduces calibration time by 68%, and exhibits strong generalization and reconstruction stability. This work establishes a new paradigm for efficient, interpretable MPI system calibration.

0 citationsRead paper

Hybrid A* Path Planning with Multi-Modal Motion Extension for Four-Wheel Steering Mobile Robots

Sep 07, 2025

To address the limited path-planning flexibility of four-wheel independent-steering (4WIS) robots in complex environments caused by single-kinematic-model abstractions, this paper proposes a multimodal-fusion Hybrid A* planning framework. The method constructs a unified four-dimensional state space to jointly represent diverse steering modes and designs a multimodal Reeds–Shepp curve set supporting forward/backward motion under differential, Ackermann, and omnidirectional kinematics. A mode-switching cost-aware heuristic function is introduced to guide search efficiently, and an intelligent terminal connection strategy enables optimal mode selection and seamless trajectory stitching. Experimental results demonstrate that the proposed approach significantly improves planning success rate and computational efficiency in narrow, dynamic environments, while enhancing motion adaptability and environmental robustness. This work establishes a scalable, multimodal planning paradigm for autonomous navigation of 4WIS platforms.

0 citationsRead paper

Recent publications

Latest Papers

FastDSAC: Enhancing Policy Plasticity via Constrained Exploration for Scalable Humanoid Locomotion

Jun 30, 2026

This work addresses the instability in value estimation and degradation of policy plasticity commonly observed in high-throughput sampling scenarios with frequent data updates. To mitigate these issues, the authors propose FastDSAC, an algorithm built upon a distributed Actor-Critic framework that models the policy using a truncated Gaussian distribution to simultaneously respect action constraints and preserve exploratory stochasticity. The method incorporates an adaptive variance modulation mechanism to enhance the accuracy of value estimation and employs implicit regularization to maintain the adaptability of the policy network. Experimental results demonstrate that FastDSAC achieves more stable training dynamics, faster convergence, and superior asymptotic performance compared to existing approaches on both the MuJoCo Playground and HumanoidBench benchmarks.

0 citationsRead paper

UAV Trajectory and Bandwidth Allocation for Efficient Data Collection in Low-Altitude Intelligent IoT: A Hierarchical DRL Approach

Apr 25, 2026

This work addresses the challenges of unknown interference, dynamic data volumes, and limited onboard computational resources in unmanned aerial vehicle (UAV)-based data collection within low-altitude intelligent Internet of Things (IoT) systems. To tackle these issues, the authors propose a lightweight hierarchical deep reinforcement learning framework, termed TBH-DDPG, which jointly optimizes data collection efficiency through coarse-grained trajectory planning at the upper level and fine-grained bandwidth allocation at the lower level. Evaluated in realistic environments featuring interference sources, time-varying data demands, and diverse obstacles, the proposed method achieves rapid convergence and low computational overhead. Compared to non-hierarchical approaches, it improves convergence speed by 44.44% and reduces computational cost by 58.05%, significantly enhancing total data throughput and overall system performance.

0 citationsRead paper

Revisiting and Expanding the IPv6 Network Periphery: Global-Scale Measurement and Security Analysis

Apr 21, 2026

This study addresses structural security risks at the IPv6 network edge—such as service exposure and routing loops—for which global-scale systematic assessment has been lacking. The authors present the first comprehensive IPv6 edge security measurement covering 73 countries, introducing a Response-Guided Prefix Selection (RGPS) strategy to efficiently scan high-value targets. They further develop a Hierarchical Large Language Model Exposure Verification (HLEV) framework to analyze unauthorized access risks. Their analysis identifies 281.9 million active IPv6 edge nodes, with a service exposure rate of 2.5%, and detects 4.5 million routing loop responses. The work also uncovers multiple security vulnerabilities in widely used LLM deployment tools, stemming from missing authentication mechanisms due to insecure default configurations.

0 citationsRead paper

Position-Prior-Guided Network for System Matrix Super-Resolution in Magnetic Particle Imaging

Nov 08, 2025

To address the time-consuming and frequently repeated system matrix (SM) calibration in magnetic particle imaging (MPI), this paper proposes a physics-informed deep super-resolution method. The core innovation lies in the first incorporation of SM’s intrinsic spatial symmetry and positional prior into the network architecture, realized via a position-guided convolutional module that jointly enforces physical constraints and data-driven learning. The method enables 2D and 3D SM reconstruction without additional hardware or measurements. Experiments demonstrate that, at identical downsampling ratios, our approach achieves ≥2.1 dB higher PSNR and ≥0.03 higher SSIM than purely data-driven baselines, reduces calibration time by 68%, and exhibits strong generalization and reconstruction stability. This work establishes a new paradigm for efficient, interpretable MPI system calibration.

0 citationsRead paper

Hybrid A* Path Planning with Multi-Modal Motion Extension for Four-Wheel Steering Mobile Robots

Sep 07, 2025

To address the limited path-planning flexibility of four-wheel independent-steering (4WIS) robots in complex environments caused by single-kinematic-model abstractions, this paper proposes a multimodal-fusion Hybrid A* planning framework. The method constructs a unified four-dimensional state space to jointly represent diverse steering modes and designs a multimodal Reeds–Shepp curve set supporting forward/backward motion under differential, Ackermann, and omnidirectional kinematics. A mode-switching cost-aware heuristic function is introduced to guide search efficiently, and an intelligent terminal connection strategy enables optimal mode selection and seamless trajectory stitching. Experimental results demonstrate that the proposed approach significantly improves planning success rate and computational efficiency in narrow, dynamic environments, while enhancing motion adaptability and environmental robustness. This work establishes a scalable, multimodal planning paradigm for autonomous navigation of 4WIS platforms.

0 citationsRead paper