Brief

last 24h

[9/9] 224 sources

Multi-source AI news clustered, deduplicated, and scored 0–100 across authority, cluster strength, headline signal, and time decay.

TOOL · arXiv cs.CV English(EN) · 1d

DiskChunGS: Large-Scale 3D Gaussian SLAM Through Chunk-Based Memory Management

Researchers have developed DiskChunGS, a novel 3D Gaussian Splatting SLAM system designed to overcome GPU memory limitations for large-scale 3D reconstructions. By employing an out-of-core approach, the system stores inactive scene parts on disk while keeping active regions in GPU memory. This method integrates with existing SLAM frameworks for pose estimation and loop closure, enabling consistent, large-scale reconstructions that have previously been constrained by hardware. AI

IMPACT Enables more extensive and detailed 3D reconstructions for robotics and AR/VR applications by overcoming hardware memory constraints.
TOOL · 雷峰网 (Leiphone) 中文(ZH) · 1w

Professor Wang Hesheng of Shanghai Jiao Tong University: From Drawing Maps to Predicting the Future, Traditional SLAM Is Stepping Out of the Static World | ICRA 2026

Researchers at Shanghai Jiao Tong University, led by Professor Hesheng Wang, are advancing Simultaneous Localization and Mapping (SLAM) beyond static environments. Their work focuses on enabling robots to navigate and understand dynamic, semantic, and deformable spaces, crucial for applications like autonomous driving and surgical robotics. The team has developed techniques for multi-modal sensor fusion, dynamic Gaussian SLAM, and modeling deformable objects, aiming to equip robots with long-term memory and reasoning capabilities for more intelligent navigation. AI

IMPACT Enables robots to navigate complex, real-world scenarios by understanding dynamic and deformable environments, pushing the boundaries of embodied AI.
TOOL · 雷峰网 (Leiphone) 中文(ZH) · 1w

ICRA 2026 Officially Opens! Rat Exoskeleton Goes Viral, VLA Route Divergence Heats Up Opening Day, 28 Papers Featured

The International Conference on Robotics and Automation (ICRA) 2026 has commenced in Vienna, featuring over 8,000 scholars. The opening day saw significant attention drawn to a rat exoskeleton for neurorehabilitation, which garnered substantial interaction. Discussions also intensified around the differing technical routes for learning in robotics, specifically reinforcement learning versus behavioral cloning, and numerous papers were presented on advancements in tactile sensing, vision-language-action (VLA) control, and simultaneous localization and mapping (SLAM). AI

IMPACT Showcases advancements in robotic perception and control, potentially accelerating the development of more capable and adaptable robots.
TOOL · arXiv cs.CV English(EN) · 3w

Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth

Researchers have introduced Depth2Pose, a new benchmark for evaluating monocular depth estimation models. This framework assesses depth quality based on the accuracy of camera pose estimation, a more practical metric for downstream tasks like visual localization and SLAM. Unlike traditional methods requiring expensive per-pixel depth data, Depth2Pose utilizes readily available camera poses, enabling evaluation in challenging environments where ground-truth depth is difficult to acquire. The accompanying D2P dataset features scenes outside the typical distribution of existing training data, highlighting potential generalization issues with current models. AI

IMPACT Introduces a new evaluation framework for depth estimation models, potentially improving their utility in real-world geometric applications.
RESEARCH · arXiv cs.CV English(EN) · 1mo · [2 sources]

DynoSLAM: Dynamic SLAM with Generative Graph Neural Networks for Real-World Social Navigation

Researchers have developed DynoSLAM, a novel Dynamic GraphSLAM architecture that integrates Graph Neural Networks (GNNs) into factor graph optimization for improved robot navigation in crowded environments. This system models pedestrian motion forecasting as a stochastic World Model, using Monte Carlo rollouts from a trained GNN to capture human interaction uncertainties. The approach embeds this uncertainty into the SLAM graph, enabling more accurate tracking and preventing optimization failures, ultimately providing a probabilistic safety envelope for collision-free robot navigation. AI

IMPACT Enhances robot navigation in dynamic, human-populated spaces by improving localization and safety prediction.
RESEARCH · arXiv cs.CV English(EN) · 1mo · [2 sources]

FreeOcc: Training-Free Embodied Open-Vocabulary Occupancy Prediction

Researchers have developed FreeOcc, a novel framework for open-vocabulary occupancy prediction that does not require any prior training or 3D annotations. This system processes monocular or RGB-D image sequences to build globally consistent occupancy maps. FreeOcc utilizes a SLAM backbone for pose estimation, a Gaussian update for dense mapping, and integrates semantics from vision-language models to achieve its predictions. AI

IMPACT Offers a training-free approach to 3D occupancy prediction, potentially reducing data requirements for robotics and AR/VR applications.
RESEARCH · arXiv cs.CV English(EN) · 1mo · [2 sources]

Holo360D: A Large-Scale Real-World Dataset with Continuous Trajectories for Advancing Panoramic 3D Reconstruction and Beyond

Researchers have introduced Holo360D, a new large-scale dataset designed to improve panoramic 3D reconstruction. This dataset features over 109,000 panoramas with registered point clouds, meshes, and camera poses, addressing limitations of existing datasets that use discrete camera locations. Holo360D provides continuous trajectories and high-completeness depth maps, collected using a 3D laser scanner and 360 camera, and enhanced with SLAM systems and a tailored post-processing pipeline. The dataset aims to establish a new benchmark for advancing feed-forward 3D reconstruction models, particularly in multi-view scenarios. AI

IMPACT Provides a new benchmark and dataset to advance panoramic 3D reconstruction models.
- arXiv
- Holo360D
RESEARCH · arXiv cs.CV English(EN) · 1mo · [2 sources]

Flow4DGS-SLAM: Optical Flow-Guided 4D Gaussian Splatting SLAM

Researchers have developed Flow4DGS-SLAM, a novel framework that enhances Simultaneous Localization and Mapping (SLAM) by integrating optical flow with 4D Gaussian Splatting. This approach aims to improve the reconstruction of both static and dynamic environments, a long-standing challenge in SLAM. The system utilizes optical flow to decompose motion, separate dynamic and static elements, and initialize camera poses, while also optimizing training speed and dynamic modeling for more accurate scene reconstruction. AI

IMPACT Improves dynamic environment reconstruction in SLAM systems, potentially benefiting applications like robotics and augmented reality.
RESEARCH · arXiv cs.CV Italiano(IT) · 1mo · [2 sources]

Robust Camera-to-Mocap Calibration and Verification for Large-Scale Multi-Camera Data Capture

Researchers have developed a new system for calibrating and verifying multi-camera setups with optical motion capture, specifically addressing challenges posed by fisheye lenses. The system enhances robustness against common errors like attachment variations and calibration drift, ensuring more reliable data for AR/VR, SLAM, and robotics applications. Experiments on Meta Quest 3 headsets demonstrated superior calibration performance and effective detection of degradation over time, with the system already integrated into production data collection pipelines. AI

IMPACT Improves data integrity for AR/VR, SLAM, and robotics, potentially enabling more robust AI training datasets.
- arXiv
- robotics
- Meta Quest 3
- AR/VR