Selected work

Representative Papers

ECO: Incremental Ego-Centric Octree Update for Point Streams

Jul 06, 2026

This work addresses the high latency and structural imbalance of conventional global octrees when processing continuous point cloud streams in real time for mobile robots. To overcome these limitations, the authors propose an Egocentric Octree (ECO) framework that dynamically confines the mapping space to the robot’s local neighborhood and employs a 3D sliding window mechanism. ECO enables efficient incremental updates through region-based classification—categorizing regions as exiting, entering, or overlapping—thereby eliminating redundant global coordinate transformations. The method maintains tree balance while preserving short-term memory of dynamic objects. Experimental results on the KITTI dataset demonstrate that ECO achieves speedups of 24.87% and 54.60% over static reconstruction and baseline incremental approaches, respectively, and accelerates downstream voxel map generation by up to 34.17%.

0 citationsRead paper

Teacher Supervision over Representation Equivalence Classes

Jul 03, 2026

This work challenges the common practice in knowledge distillation of naively matching a teacher model’s absolute feature representations, which overlooks the fact that such representations are only equivalent up to orthogonal transformations and isotropic scaling. From a geometric perspective, the paper proposes a new paradigm centered on representation equivalence classes: the student should instead learn class-invariant structures of the teacher’s representations—such as Gram matrices, centered kernel alignment (CKA), or principal subspaces—or leverage coordinate alignment for effective supervision. This framework unifies feature matching, relational distillation, and grafting approaches, revealing that logit-level matching is ultimately key to capability transfer. Experiments on Qwen2.5 and Llama-3.1 demonstrate that high CKA similarity alone is insufficient for performance recovery, while successful grafting hinges on boundary overlap in the training data coverage, thereby validating the proposed theory.

0 citationsRead paper

Comparative Study on Agility, Efficiency, and Impact Absorption of Bipedal Robots with Active Toes

Jun 17, 2026

This study addresses the challenge that existing bipedal robots fail to replicate the critical role of human toes in agility, energy efficiency, and impact absorption, and lack quantitative validation of active toe benefits. To bridge this gap, the authors develop a 14-degree-of-freedom anthropomorphic bipedal robot platform featuring a lightweight, high-torque, and robust active toe mechanism. Using a unified reinforcement learning training pipeline in a high-fidelity simulation environment, they present the first quantitative assessment of active toes’ comprehensive benefits under identical conditions. Experimental results demonstrate that, at a walking speed of 1.33 m/s, the active toe configuration reduces the cost of transport (CoT) by 17.5%, decreases heel-strike ground reaction forces by 5.0%, and lowers both the average and peak path deviations by 25.0% and 34.0%, respectively, compared to a no-toe configuration.

0 citationsRead paper

Hierarchical vs. Flat Iteration in Shared-Weight Transformers

Apr 15, 2026

This study investigates whether hierarchically weight-shared recurrent architectures can match the representational capacity of conventionally stacked, independent Transformer layers. To this end, we propose HRM-LM, a model employing a dual-speed recurrence mechanism: a fast module updates representations at every step to capture local structure, while a slow module performs global compression every T steps, yielding an effective computational depth equivalent to N×T layers under shared parameters. Using an ablation framework based on the Universal Transformer and conducting five independent trials at a 1.2B-parameter scale, we consistently observe that this hierarchical recurrent architecture significantly underperforms standard Transformers in representation quality. Our work systematically reveals a fundamental gap in expressive power between hierarchical iterative designs and flatly stacked Transformer layers.

0 citationsRead paper

CV-HoloSR: Hologram to hologram super-resolution through volume-upsampling three-dimensional scenes

Apr 11, 2026

This work addresses the severe quadratic depth distortion introduced by existing holographic super-resolution methods during volumetric upsampling, which compromises three-dimensional focus accuracy. To overcome this limitation, the authors propose CV-HoloSR, a novel framework that leverages a complex-valued residual dense network and a depth-aware perceptual reconstruction loss to achieve, for the first time, physically consistent linear depth scaling in volumetric upsampling, thereby effectively recovering high-frequency interference details. The study introduces the first 4K holographic dataset tailored for large depth ranges and devises a complex-valued LoRA fine-tuning strategy that enables efficient adaptation to new depth and display configurations with only 200 samples. Experiments demonstrate a 32% improvement in LPIPS (reaching 0.2001), a 75% reduction in training time (from 22.5 to 5.2 hours), and successful generalization to unseen depth ranges and novel display setups.

0 citationsRead paper

Recent publications

Latest Papers

ECO: Incremental Ego-Centric Octree Update for Point Streams

Jul 06, 2026

This work addresses the high latency and structural imbalance of conventional global octrees when processing continuous point cloud streams in real time for mobile robots. To overcome these limitations, the authors propose an Egocentric Octree (ECO) framework that dynamically confines the mapping space to the robot’s local neighborhood and employs a 3D sliding window mechanism. ECO enables efficient incremental updates through region-based classification—categorizing regions as exiting, entering, or overlapping—thereby eliminating redundant global coordinate transformations. The method maintains tree balance while preserving short-term memory of dynamic objects. Experimental results on the KITTI dataset demonstrate that ECO achieves speedups of 24.87% and 54.60% over static reconstruction and baseline incremental approaches, respectively, and accelerates downstream voxel map generation by up to 34.17%.

0 citationsRead paper

Teacher Supervision over Representation Equivalence Classes

Jul 03, 2026

This work challenges the common practice in knowledge distillation of naively matching a teacher model’s absolute feature representations, which overlooks the fact that such representations are only equivalent up to orthogonal transformations and isotropic scaling. From a geometric perspective, the paper proposes a new paradigm centered on representation equivalence classes: the student should instead learn class-invariant structures of the teacher’s representations—such as Gram matrices, centered kernel alignment (CKA), or principal subspaces—or leverage coordinate alignment for effective supervision. This framework unifies feature matching, relational distillation, and grafting approaches, revealing that logit-level matching is ultimately key to capability transfer. Experiments on Qwen2.5 and Llama-3.1 demonstrate that high CKA similarity alone is insufficient for performance recovery, while successful grafting hinges on boundary overlap in the training data coverage, thereby validating the proposed theory.

0 citationsRead paper

Comparative Study on Agility, Efficiency, and Impact Absorption of Bipedal Robots with Active Toes

Jun 17, 2026

This study addresses the challenge that existing bipedal robots fail to replicate the critical role of human toes in agility, energy efficiency, and impact absorption, and lack quantitative validation of active toe benefits. To bridge this gap, the authors develop a 14-degree-of-freedom anthropomorphic bipedal robot platform featuring a lightweight, high-torque, and robust active toe mechanism. Using a unified reinforcement learning training pipeline in a high-fidelity simulation environment, they present the first quantitative assessment of active toes’ comprehensive benefits under identical conditions. Experimental results demonstrate that, at a walking speed of 1.33 m/s, the active toe configuration reduces the cost of transport (CoT) by 17.5%, decreases heel-strike ground reaction forces by 5.0%, and lowers both the average and peak path deviations by 25.0% and 34.0%, respectively, compared to a no-toe configuration.

0 citationsRead paper

Hierarchical vs. Flat Iteration in Shared-Weight Transformers

Apr 15, 2026

This study investigates whether hierarchically weight-shared recurrent architectures can match the representational capacity of conventionally stacked, independent Transformer layers. To this end, we propose HRM-LM, a model employing a dual-speed recurrence mechanism: a fast module updates representations at every step to capture local structure, while a slow module performs global compression every T steps, yielding an effective computational depth equivalent to N×T layers under shared parameters. Using an ablation framework based on the Universal Transformer and conducting five independent trials at a 1.2B-parameter scale, we consistently observe that this hierarchical recurrent architecture significantly underperforms standard Transformers in representation quality. Our work systematically reveals a fundamental gap in expressive power between hierarchical iterative designs and flatly stacked Transformer layers.

0 citationsRead paper

CV-HoloSR: Hologram to hologram super-resolution through volume-upsampling three-dimensional scenes

Apr 11, 2026

This work addresses the severe quadratic depth distortion introduced by existing holographic super-resolution methods during volumetric upsampling, which compromises three-dimensional focus accuracy. To overcome this limitation, the authors propose CV-HoloSR, a novel framework that leverages a complex-valued residual dense network and a depth-aware perceptual reconstruction loss to achieve, for the first time, physically consistent linear depth scaling in volumetric upsampling, thereby effectively recovering high-frequency interference details. The study introduces the first 4K holographic dataset tailored for large depth ranges and devises a complex-valued LoRA fine-tuning strategy that enables efficient adaptation to new depth and display configurations with only 200 samples. Experiments demonstrate a 32% improvement in LPIPS (reaching 0.2001), a 75% reduction in training time (from 22.5 to 5.2 hours), and successful generalization to unseen depth ranges and novel display setups.

0 citationsRead paper