返回首页

CV

共 68 篇论文

cs.CV自动驾驶CV

Reconstruction Matters: Learning Geometry-Aligned BEV Representation through 3D Gaussian Splatting

Bird's-Eye-View (BEV) perception is fundamental for autonomous driving, providing a unified spatial representation that fuses surrounding-view images for downstream tasks like semantic segmentation, 3D object detection, and motion prediction. Current end-to-end BEV frameworks treat image-to-BEV transformation as a black box, lacking explicit 3D geometric understanding and often yielding suboptimal performance. This paper introduces Splat2BEV, a Gaussian Splatting-assisted framework that learns BEV feature representations combining semantic richness with geometric precision. By leveraging 3D Gaussian Splatting for reconstruction, the method explicitly incorporates 3D geometry awareness into the BEV perception pipeline, addressing the limitations of traditional end-to-end approaches.

Splat2BEV Authors
about 1 month ago
arXiv 2603.19193v1
cs.CVCVTransformer

Sparse Autoencoders Reveal Interpretable and Steerable Features in VLA Models

Vision-Language-Action (VLA) models have emerged as a promising approach for general-purpose robot manipulation, combining visual perception, language understanding, and action planning in a unified framework. This work applies mechanistic interpretability techniques using Sparse Autoencoders (SAEs) to analyze hidden layer activations in VLA models, revealing sparse dictionary features that provide interpretable bases for model computation. The research discovers that most SAE features correspond to memorized sequences from training demonstrations, while some features represent interpretable, generalizable motion primitives and semantic properties. This analysis offers insights into VLA model generalizability and provides a framework for steering model behavior through identified interpretable features, advancing the understanding of embodied AI systems for robot manipulation tasks.

Anonymous Authors
about 1 month ago
arXiv 2603.19183v1
cs.CV端到端CV

ATG-MoE: Autoregressive trajectory generation with mixture-of-experts for assembly skill learning

This paper presents ATG-MoE, an end-to-end autoregressive trajectory generation method with mixture-of-experts architecture for robot assembly skill learning from demonstration. The method processes multi-modal inputs including RGB-D observations, natural language instructions, and robot proprioception to generate manipulation trajectories in a closed-loop manner. It incorporates multi-modal feature fusion for comprehensive scene and task understanding, autoregressive sequence modeling for temporally coherent trajectory generation, and a mixture-of-experts architecture enabling unified multi-skill learning. The approach addresses challenges in flexible manufacturing where robot systems must adapt to changing tasks, objects, and environments without labor-intensive traditional programming.

ATG-MoE Authors
about 1 month ago
arXiv 2603.19029v1
cs.CVCV3D检测

GHOST: Fast Category-agnostic Hand-Object Interaction Reconstruction from RGB Videos using Gaussian Splatting

This paper presents GHOST (Gaussian Hand-Object Splatting), a fast category-agnostic framework for reconstructing dynamic hand-object interactions from monocular RGB videos. The method represents both hands and objects as dense, view-consistent Gaussian discs to achieve complete 3D reconstructions. Three key innovations are introduced: a geometric-prior retrieval and consistency loss for completing occluded object regions, grasp-aware alignment for refining hand translations and object scale to ensure realistic contact, and a hand-aware background loss that prevents penalizing hand-occluded object regions. The framework enables physically consistent and animatable reconstructions while running an order of magnitude faster than existing methods, with applications in AR/VR, robotics, and embodied AI.

Anonymous Authors
about 1 month ago
arXiv 2603.18912v1
cs.CVCVTransformer

MultihopSpatial: Multi-hop Compositional Spatial Reasoning Benchmark for Vision-Language Model

Spatial reasoning is foundational for Vision-Language Models deployed as Vision-Language-Action agents in physical environments. This paper introduces MultihopSpatial, a benchmark designed for multi-hop and compositional spatial reasoning with complex queries across diverse spatial perspectives. The work proposes Acc@50IoU, a metric that simultaneously evaluates reasoning and visual grounding by requiring both answer selection and precise bounding box prediction. MultihopSpatial-Train provides a large-scale training corpus to foster spatial intelligence. Extensive evaluation of 37 state-of-the-art VLMs yields key insights into the capabilities and limitations of current models for robust VLA deployment in embodied AI scenarios.

MultihopSpatial Team
about 1 month ago
arXiv 2603.18892v1
cs.CVCV3D检测

V-Dreamer: Automating Robotic Simulation and Trajectory Synthesis via Video Generation Priors

This paper presents V-Dreamer, a fully automated framework for generating simulation-ready manipulation environments and executable robot trajectories from natural language instructions. The system leverages large language models and 3D generative models to construct physically grounded 3D scenes validated by geometric constraints for stable, collision-free layouts. Video generation models serve as rich motion priors for behavior synthesis, which are then mapped to executable robot trajectories through a Sim-to-Gen visual-kinematic alignment module using CoTracker3 and VGGT. This approach enables high visual diversity and physical fidelity for robotic manipulation training.

Anonymous
about 1 month ago
arXiv 2603.18811v1
cs.CVCV目标检测

Robotic Agentic Platform for Intelligent Electric Vehicle Disassembly

This paper presents RAPID, a Robotic Agentic Platform for Intelligent Disassembly designed to address the challenge of scalable EV battery recycling. The system features a gantry-mounted industrial manipulator with RGB-D perception capabilities and an automated nut-running tool for fastener removal on full-scale EV battery packs. An open-vocabulary object detection pipeline achieves 0.9757 mAP50 for reliable identification of screws, nuts, busbars, and components. The research experimentally evaluates three one-shot fastener removal strategies: taught-in poses (97% success, 24 min), one-shot vision execution (57%, 29 min), and visual servoing (83%, 36 min), providing a comprehensive comparison of success rates and disassembly times for battery top cover fasteners.

RAPID Research Team
about 1 month ago
arXiv 2603.18520v1
cs.CVCV3D检测

MGSO: Monocular Real-time Photometric SLAM with Efficient 3D Gaussian Splatting

This paper presents MGSO (Monocular Gaussian Splatting Optimization), a novel real-time SLAM system that integrates photometric SLAM with efficient 3D Gaussian Splatting for dense 3D reconstruction. The proposed approach leverages photometric SLAM to generate dense structured point clouds that accelerate 3D Gaussian initialization and optimization. By producing more efficient maps with fewer Gaussians while maintaining reconstruction quality, the system achieves an excellent balance between quality, memory efficiency, and speed. Experiments demonstrate that MGSO outperforms state-of-the-art 3DGS-based SLAM systems, making it particularly suitable for real-time dense mapping on resource-limited devices.

MGSO Authors
over 1 year ago
arXiv 2409.13055v3
上一页第 4 / 4 页