Selected work

Representative Papers

ResemBrick: Brick Reconstruction from Photographs with Perceptual Fidelity and Buildability

Aug 10, 2026

This work addresses the challenge of generating visually realistic and physically constructible colored brick models from a small set of casually captured photographs, under constraints of physical assemblability and a fixed discrete voxel budget. To this end, the authors propose an end-to-end method that formulates voxel discretization as an optimizable resource allocation problem for the first time. Their approach employs a resolution-conditioned network for budget-aware voxel occupancy prediction, followed by a support-aware greedy placement strategy and a deterministic repair algorithm to guarantee structural stability and eliminate unsupported bricks. Under identical voxel budgets, the method achieves significantly higher perceptual fidelity compared to prior approaches and produces zero unsupported or unstable components on an unfiltered test set.

0 citationsRead paper

Dependency-Aware Reliability Allocation for Open-Vocabulary Scene-Graph Semantic Packets over Latency- and Energy-Constrained Wireless Visual Uplinks

Aug 08, 2026

This work addresses the challenge of preserving semantic interpretability in wireless visual uplink transmissions under stringent delay and energy constraints, where open-vocabulary scene graph semantic packets suffer from degraded reliability due to vocabulary expansion and triplet dependencies. To tackle this, the authors propose constructing a semantic packet dependency graph and jointly optimizing modulation and coding schemes, transmit power, and the maximum number of HARQ retransmissions to minimize dependency-aware semantic distortion while satisfying frame-level latency and energy budgets. The approach introduces a shared downstream enablement value mechanism and a semantic dependency Lagrangian block method, enabling precise modeling of dependency structures and yielding near-global optimal solutions within a limited candidate set. Experiments demonstrate that the method achieves global optimality in 28 out of 30 test instances, with an average relative error of merely 0.052%, significantly outperforming baseline approaches by reducing semantic failure rates and enhancing critical query success rates.

0 citationsRead paper

Design and Control of the "QuadBoat": A Quadruped Surface Vehicle for Drowning Rescue

Jul 15, 2026

This work proposes QuadBoat, a bio-inspired quadrupedal unmanned surface vessel designed to overcome the limitations of existing water rescue robots in rapidly and accurately retrieving drowning victims. By introducing a quadrupedal configuration to surface rescue for the first time, QuadBoat achieves high adaptability and agile maneuverability through active posture regulation. The system integrates inverse kinematics-based motion control, a cascaded MPC-PID controller, and a vision-based tracking module to enable robust target detection, pursuit, and retrieval. Experimental results demonstrate that QuadBoat attains exceptional trajectory tracking accuracy, superior agility, and effective drowning victim recovery capabilities in both indoor and outdoor environments, significantly enhancing the efficiency and flexibility of aquatic rescue operations.

0 citationsRead paper

Self-Improving is Often Sudden: Enlightenment-style Finetuning for Large-Scale Models

Jul 14, 2026

Inspired by the human phenomenon of “insight,” this work proposes Enlightenment, a training-free post-tuning paradigm designed to elicit abrupt, emergent capability improvements in large-scale models. Moving beyond existing training-free approaches that solely adjust attention weights, Enlightenment introduces an architecture-aware shortcut rewiring mechanism—specifically, attention head mixing shortcuts for large language models (LLMs) and scalar modulation of residual connections for vision-language models (VLMs). Experimental results demonstrate that Enlightenment consistently achieves significant performance gains across diverse models and benchmarks, effectively unlocking latent capabilities embedded in pretrained architectures without additional training.

0 citationsRead paper

Explicit Stair Geometry Conditioning for Robust Humanoid Locomotion

May 10, 2026

This work addresses the challenges humanoids face when climbing stairs in real-world environments—namely abrupt geometric transitions, height sensitivity, and perceptual uncertainty—which limit the generalization of existing methods. The authors propose a reinforcement learning strategy conditioned on explicit geometric parameters, directly extracting interpretable features such as step height, depth, and yaw angle to modulate a PPO-based locomotion controller. By eschewing implicit terrain representations in favor of compact, interpretable geometric priors, the approach significantly enhances generalization across diverse stair structures and enables proactive gait adaptation. Simulations demonstrate effective out-of-distribution generalization to unseen step heights, and real-world experiments on the Unitree G1 robot achieve robust indoor and outdoor stair climbing, including uninterrupted ascension of 33 consecutive steps outdoors without failure.

0 citationsRead paper

Recent publications

Latest Papers

ResemBrick: Brick Reconstruction from Photographs with Perceptual Fidelity and Buildability

Aug 10, 2026

This work addresses the challenge of generating visually realistic and physically constructible colored brick models from a small set of casually captured photographs, under constraints of physical assemblability and a fixed discrete voxel budget. To this end, the authors propose an end-to-end method that formulates voxel discretization as an optimizable resource allocation problem for the first time. Their approach employs a resolution-conditioned network for budget-aware voxel occupancy prediction, followed by a support-aware greedy placement strategy and a deterministic repair algorithm to guarantee structural stability and eliminate unsupported bricks. Under identical voxel budgets, the method achieves significantly higher perceptual fidelity compared to prior approaches and produces zero unsupported or unstable components on an unfiltered test set.

0 citationsRead paper

Dependency-Aware Reliability Allocation for Open-Vocabulary Scene-Graph Semantic Packets over Latency- and Energy-Constrained Wireless Visual Uplinks

Aug 08, 2026

This work addresses the challenge of preserving semantic interpretability in wireless visual uplink transmissions under stringent delay and energy constraints, where open-vocabulary scene graph semantic packets suffer from degraded reliability due to vocabulary expansion and triplet dependencies. To tackle this, the authors propose constructing a semantic packet dependency graph and jointly optimizing modulation and coding schemes, transmit power, and the maximum number of HARQ retransmissions to minimize dependency-aware semantic distortion while satisfying frame-level latency and energy budgets. The approach introduces a shared downstream enablement value mechanism and a semantic dependency Lagrangian block method, enabling precise modeling of dependency structures and yielding near-global optimal solutions within a limited candidate set. Experiments demonstrate that the method achieves global optimality in 28 out of 30 test instances, with an average relative error of merely 0.052%, significantly outperforming baseline approaches by reducing semantic failure rates and enhancing critical query success rates.

0 citationsRead paper

Design and Control of the "QuadBoat": A Quadruped Surface Vehicle for Drowning Rescue

Jul 15, 2026

This work proposes QuadBoat, a bio-inspired quadrupedal unmanned surface vessel designed to overcome the limitations of existing water rescue robots in rapidly and accurately retrieving drowning victims. By introducing a quadrupedal configuration to surface rescue for the first time, QuadBoat achieves high adaptability and agile maneuverability through active posture regulation. The system integrates inverse kinematics-based motion control, a cascaded MPC-PID controller, and a vision-based tracking module to enable robust target detection, pursuit, and retrieval. Experimental results demonstrate that QuadBoat attains exceptional trajectory tracking accuracy, superior agility, and effective drowning victim recovery capabilities in both indoor and outdoor environments, significantly enhancing the efficiency and flexibility of aquatic rescue operations.

0 citationsRead paper

Self-Improving is Often Sudden: Enlightenment-style Finetuning for Large-Scale Models

Jul 14, 2026

Inspired by the human phenomenon of “insight,” this work proposes Enlightenment, a training-free post-tuning paradigm designed to elicit abrupt, emergent capability improvements in large-scale models. Moving beyond existing training-free approaches that solely adjust attention weights, Enlightenment introduces an architecture-aware shortcut rewiring mechanism—specifically, attention head mixing shortcuts for large language models (LLMs) and scalar modulation of residual connections for vision-language models (VLMs). Experimental results demonstrate that Enlightenment consistently achieves significant performance gains across diverse models and benchmarks, effectively unlocking latent capabilities embedded in pretrained architectures without additional training.

0 citationsRead paper

Explicit Stair Geometry Conditioning for Robust Humanoid Locomotion

May 10, 2026

This work addresses the challenges humanoids face when climbing stairs in real-world environments—namely abrupt geometric transitions, height sensitivity, and perceptual uncertainty—which limit the generalization of existing methods. The authors propose a reinforcement learning strategy conditioned on explicit geometric parameters, directly extracting interpretable features such as step height, depth, and yaw angle to modulate a PPO-based locomotion controller. By eschewing implicit terrain representations in favor of compact, interpretable geometric priors, the approach significantly enhances generalization across diverse stair structures and enables proactive gait adaptation. Simulations demonstrate effective out-of-distribution generalization to unseen step heights, and real-world experiments on the Unitree G1 robot achieve robust indoor and outdoor stair climbing, including uninterrupted ascension of 33 consecutive steps outdoors without failure.

0 citationsRead paper