Apple's research team published an arXiv paper detailing HALO, a humanoid robot autonomy stack that layers a VLM task orchestrator, a DiT motion generator, and a reinforcement-learning tracker, deployed on a Unitree G1 with an onboard Mac mini for inference. Across 360 single-task trials in six previously unmapped real environments, HALO averaged a 91% success rate, though long-horizon composite tasks fell to 70% and the hardest combination to 50%. The team also open-sourced the data generation, training, and deployment code.
Vision-Language Model
视觉-语言模型
Multimodal architecture jointly processing visual and textual info
Curated Intelligence
Amazon describes a multi-model fusion approach built on SageMaker AI to evaluate the holistic quality of A+ Content, where text, images, and video are judged together rather than asset by asset. It is a vendor-bound engineering practice write-up aimed at heterogeneous multi-module content, without disclosed model parameters or reproducible benchmark data.
Tencent has unveiled Hy4, its new flagship model that for the first time understands images and can operate software autonomously, with targeted optimization for coding, finance and office work. Tencent claims second place among Chinese models, but independent tests show it trails top systems like Claude Opus 5.5 on hard math and coding tasks; model weights will be released later.
Google Research has open-sourced Agile Deliberation, a framework that lets users define fuzzy 'gray area' concepts before an AI classifier learns them. The system splits a concept into in-scope and out-of-scope subconcepts, surfaces borderline samples for user labeling, and has Gemini rewrite the definition as a VLM prompt. Across 18 live sessions it beat zero-shot Gemini 2.5 Flash by 10.5 F1 points on average, with code released under Apache 2.0.
A team from Xiamen University and Nanyang Technological University proposes RoTA-Bench, an interactive 3D-to-2D projection alignment benchmark that tests whether vision-language models can turn 2D observations into 3D understanding precise enough to drive rotation actions. It fills the gap between static spatial QA without action feedback and general interactive benchmarks that do not isolate geometric pose reasoning, targeting the perception-control link in embodied agents.
Lightweight vision models Moondream 3 Preview and 3.1 are now available on OpenRouter, served by PrimeIntellect. This update enables users to perform object detection and pixel-level pointing directly within a standard chat interface by attaching an image and using simple commands like /detect or /point, eliminating the need for complex API configurations.
CLIPSeg leverages CLIP's vision-language alignment to achieve zero-shot image segmentation without task-specific annotations. The architecture maps text prompts directly to pixel-level predictions, significantly reducing reliance on large-scale manual labeling.