SpatialBrain fixes the vision-language backbone and adds memory of reasoning decisions to improve spatial tasks; it fuses current observations into spatial relations, uses reusable tools, and links task conditions to experienced strategies, achieving strong results across five benchmarks.
MJev-Spatial is a single-forward-pass multimodal model for spatial reasoning, treating candidates as classes and using pretrained projections; it trains with joint coarse-to-fine supervision and teacher forcing, achieves high accuracy and faster throughput on spatial benchmarks, built on Qwen3.5-4B.
SonicWorld combines physical, acoustic, and semantic cues to model moving sound sources and reason spatially; it introduces Sonic-Bench for evaluation and reports improved localization accuracy and azimuth precision over a prior baseline, with strong zero-shot performance on LibriSpeech.
Databricks introduces Lakehouse//RT for real-time geospatial analytics, enabling Spatial SQL with native GEOMETRY types and 100+ functions. It offers significant speedups (up to 38x under high concurrency) versus traditional data warehouses and maintains portability with no code changes. The solution targets diverse industries and provides a beta with steps to get started.
Google Research outlines a Geospatial Reasoning agent that integrates Earth AI foundation models with planet-scale data to perform modular, explainable spatial analysis beyond traditional LLMs; BYOD support and cloud-native data access enable enterprise workflows, with early results showing improved health outcome predictions and higher spatial resolution.
Spatial Harness grounds vision-language model spatial reasoning in executable scene states, using a physics engine to test viewpoint changes and provide measurements and rendered views. On MindCube, it improved Qwen3.5-27B accuracy from 52.70% to 71.14%, and reached 98.25% with annotated object and camera poses.
Translated title. Original: 私は Gemini ですが、 GeoAI と LLM / 生成AI は特段関係がなさそうです｜ゆいせき
The article argues that GeoAI traditionally means spatially explicit AI tied to geographic data, predating the current LLM surge. It distinguishes between AI for GIS (prompt-driven automation) and GIS for AI (true spatial understanding), noting that LLMs alone cannot reliably perform spatial reasoning and can produce spatial hallucinations. It highlights historical and regional differences, with Japan leaning toward tooling for automation, while calls for a robust spatiotemporal computation backbone integrated with GIS standards. The piece urges developing true spatial intelligence rather than relying solely on LLM prompts.
FPV-SLAM is presented as a leading SLAM system for egocentric data, achieving first place in the monocular-inertial track of ETH Zurich's LaMAria city-scale benchmark.
Meta released InteractiveWorld, a research benchmark to test AI assistants in first-person tasks within a simulated environment. It features 120 tasks across six categories, uses a human participant model for evaluation, and emphasizes practical usability and data/privacy constraints over mere task completion.
Translated title. Original: Physical AI: come ALBA Robot porta l'AI nel mondo reale - Economyup
ALBA Robot developed autonomous mobility vehicles from a wheelchair-access problem into a platform for airports, hospitals, and other public spaces. Its vehicles also gather operational data, while centralized fleet management and remote updates help run them across complex sites.