Generation of Vectorized Maps Beyond Vehicle View
本文针对自动驾驶中车辆视野范围有限的问题，提出了一种名为BeyondFormer的新方法，用于生成超出视野范围的矢量化地图，以提高安全规划。
0 citationsRead paper
Selected work
Representative Papers
Multi-Robot Bearing-only Pose Estimation via Angle Rigidity
This work addresses the problem of pose estimation for multi-robot systems in three-dimensional space when no prior orientation information is available. The paper proposes a distributed bearing-only estimation algorithm that relies solely on bearing measurements and their time derivatives expressed in each robot’s body-fixed frame. By introducing angular rigidity—a novel relaxation of the conventional bearing rigidity condition—the method simultaneously estimates both positions and orientations in SO(3) without requiring any known initial headings. This constitutes the first distributed bearing-only pose estimation approach in 3D that operates without prior orientation knowledge. Theoretical analysis establishes local exponential stability of the estimation error dynamics, and simulation results demonstrate the algorithm’s effectiveness and practicality under realistic conditions.
0 citationsRead paper
Rethinking the Illusion of Thinking
This work addresses the fundamental debate over whether large reasoning models (LRMs) possess genuine reasoning capability, specifically distinguishing whether their failures on complex tasks—such as the Tower of Hanoi and river-crossing puzzles—stem from output-length constraints or intrinsic cognitive limitations. Method: We propose three key methodological innovations: (i) incremental stepwise prompting, (ii) multi-agent collaborative reasoning, and (iii) a controllable benchmarking framework, complemented by a fine-grained ablation analysis to precisely isolate failure sources. Contribution/Results: Empirical evaluation reveals that LRMs encounter a sharp cognitive bottleneck at ~8-disk Tower of Hanoi instances, yet solve river-crossing problems with >100 entity pairs reliably. These findings demonstrate that LRMs function as stochastic search-based reasoners—not mere “hallucinatory” systems—and provide both a novel evaluation paradigm and empirical grounding for assessing LRM reasoning competence.
0 citationsRead paper
Recent publications
Latest Papers
Generation of Vectorized Maps Beyond Vehicle View
本文针对自动驾驶中车辆视野范围有限的问题，提出了一种名为BeyondFormer的新方法，用于生成超出视野范围的矢量化地图，以提高安全规划。
0 citationsRead paper
Multi-Robot Bearing-only Pose Estimation via Angle Rigidity
This work addresses the problem of pose estimation for multi-robot systems in three-dimensional space when no prior orientation information is available. The paper proposes a distributed bearing-only estimation algorithm that relies solely on bearing measurements and their time derivatives expressed in each robot’s body-fixed frame. By introducing angular rigidity—a novel relaxation of the conventional bearing rigidity condition—the method simultaneously estimates both positions and orientations in SO(3) without requiring any known initial headings. This constitutes the first distributed bearing-only pose estimation approach in 3D that operates without prior orientation knowledge. Theoretical analysis establishes local exponential stability of the estimation error dynamics, and simulation results demonstrate the algorithm’s effectiveness and practicality under realistic conditions.
0 citationsRead paper
Rethinking the Illusion of Thinking
This work addresses the fundamental debate over whether large reasoning models (LRMs) possess genuine reasoning capability, specifically distinguishing whether their failures on complex tasks—such as the Tower of Hanoi and river-crossing puzzles—stem from output-length constraints or intrinsic cognitive limitations. Method: We propose three key methodological innovations: (i) incremental stepwise prompting, (ii) multi-agent collaborative reasoning, and (iii) a controllable benchmarking framework, complemented by a fine-grained ablation analysis to precisely isolate failure sources. Contribution/Results: Empirical evaluation reveals that LRMs encounter a sharp cognitive bottleneck at ~8-disk Tower of Hanoi instances, yet solve river-crossing problems with >100 entity pairs reliably. These findings demonstrate that LRMs function as stochastic search-based reasoners—not mere “hallucinatory” systems—and provide both a novel evaluation paradigm and empirical grounding for assessing LRM reasoning competence.
0 citationsRead paper