Vision Language Models

Apple has released LensVLM-9B, an artificial intelligence model designed to read and understand compressed documents. The release suggests Apple is expanding its work in document-focused vision-language systems, and details about the model's capabilities and availability are now circulating among AI watchers. Further information on benchmarks, licensing and intended use was not immediately available.

NVIDIA has introduced Nemotron 3.5 Super VL, a new vision-language model aimed at powering advanced AI agents. The release highlights the company's continued push into agentic AI, offering multimodal capabilities for developers building autonomous systems. Details on benchmarks, licensing and availability are still limited, but the announcement has drawn attention across the AI developer community.

Researchers have introduced Spa-Bench, a real-robot benchmark designed to measure spatial grounding in vision-language-action models. Early results show that evaluated robot policies perform noticeably worse when tasks move from in-distribution to out-of-distribution spatial configurations, exposing a key weakness in current systems. The paper, released on arXiv's robotics section, gives the field a standardised way to compare how well robots understand spatial relationships in real-world settings.

Researchers have presented CoPlanVLM, a new method described in a paper on arXiv that uses Vision Language Models for coordinated path planning across a team of robots. The work, posted in the robotics category, applies large multimodal models to help multiple robots plan and coordinate their routes together rather than acting independently.

Industry discussion is focusing on robotic dexterity as the main obstacle holding back physical AI. While large AI models have advanced rapidly in language and vision, robots still struggle with fine motor tasks like grasping and manipulating unfamiliar objects, which limits deployment in factories, warehouses and homes. Analysts argue that solving manipulation, not intelligence alone, is the key challenge for humanoid and industrial robotics going forward.

A new arXiv paper in robotics argues that pretrained vision-language models can act directly as manipulation controllers when grounded in an explicit action space, rather than serving only as perception or planning helpers. The approach suggests a path to reusing large pretrained models as the core 'brain' of robotic arms, a step researchers say could cut the cost of training robot-specific policies from scratch.

Researchers have presented a RAG-enhanced multi-agent framework combining large language models and vision-language models for structural health monitoring of a historic masonry building. The system is designed to be auditable and region-aware, documenting how damage assessments are reached across different parts of the structure. The work, published in Nature's portfolio, points to growing use of generative AI in civil engineering inspection and heritage conservation.

A new survey paper examines how foundation models are being used for affordance learning in robotics, tracing the pipeline from perception through to robotic manipulation. The work reviews how large pre-trained models help robots understand what actions objects afford, consolidating recent progress across vision, language and control. Researchers say the field is moving quickly as general-purpose models are adapted for physical tasks.

NVIDIA has published a technical blog introducing Nemotron 3.5 Super VL, a vision-language model designed to power specialized AI agents with multimodal reasoning. The release is aimed at developers building applications that combine text and image understanding, extending NVIDIA's Nemotron family of open models for enterprise and agentic use cases.

Researchers have proposed FastOPD, a new framework that compresses large foundation vision-language-action models into lightweight versions suitable for real-world robot deployment. The approach uses efficient on-policy distillation, allowing small models to learn directly from a big model's behavior. The work was released as a paper on arXiv in the robotics category, and early discussion is focused on whether it makes advanced robot policies cheaper and easier to run on hardware.

A new paper on arXiv introduces SocialVLA, a vision-language-action model that reads spontaneous human reactions and converts them into runtime intervention signals for robot manipulation. The aim is to let robots respond to natural human cues, like flinches or gestures, while working alongside people. Details of performance and real-world testing remain in the paper itself.

Researchers have introduced a new benchmark designed to measure how well robots powered by vision-language models retain and use memory across tasks. The tool aims to standardize evaluation of robotic memory, a key weakness in current systems, and could help developers compare progress and push robots toward more reliable real-world behaviour.