SenseTime, PHANES AI and openJiuwen Push AI Agents Into Cities, Robots and Offices

SenseTime Dax and the Shanghai Municipal Planning and Natural Resources Bureau upgraded the Yunyu Xingkong large model, described as China's first industry foundation model for planning and natural resources. The model is positioned as the intelligent base and hub for the quantum-city model system. It has 13,000 high-quality question-answer pairs and 1.5 billion tokens of corpus, covering planning and design, surveying and mapping, geography and natural resources. The upgrade adds an object-relation-action framework and links policy and business language, spatial units such as buildings, parcels and roads, and multi-source data including current conditions, plans, monitoring and remote sensing. It also strengthens map-related reasoning across retrieval, recognition, interpretation and generation. The bureau released the first batch of quantum-city government agents, including the flood-control agent Yuzhi Heluo, Better Life in Your Palm and Shanghai-Style Jiangnan Tour. Three agents, Yuzhi Heluo, Dalue Zhiyan and Xianggui Xiaogongdan, were developed and delivered by SenseTime Dax for flood-control dispatch, spatial insight and planning updates. Yuzhi Heluo uses weather, water levels and drainage flow data for district- and time-specific dispatch. It supports 48-hour and even 2-hour advance analysis and real-time simulation, balancing storage across river networks, pipe networks and blocks, according to the report. After deployment, spatial computing, dynamic simulation and intelligent dispatch efficiency rose 65%, 70% and 80%, respectively. Dalue Zhiyan aggregates 18 years of current conditions, planning, ownership and approval data and builds a digital twin base map covering buildings, farmland, forests, water and wetlands; users can query it in natural language. Xianggui Xiaogongdan supports urban renewal by handling enterprise requests, co-creation and plan review. SenseTime Dax said it used an FDE co-creation model, with business experts and research teams working on the front line. Future plans include a city-level macro world model, city agent-group collaboration and an open-source city ecosystem.

PHANES AI, founded by Yang Shuo, launched SeePhys Hub with Tianlong Mobile. Yang, born in 1998, is a tenured professor and doctoral supervisor at Harbin Institute of Technology, Shenzhen, and was described as China's youngest AI professor and doctoral supervisor. SeePhys Hub combines a four-eye headband, two wristbands, three waist-and-leg motion trackers and a tactile glove. It uses sparse tracking nodes, inertial measurement units and ultra-wideband signals, plus visual data from a 360-degree headband camera and wrist cameras, to reconstruct full-body pose and collect hand contact pressure when needed. Synchronization accuracy among the headset camera, inertial measurement unit and ultra-wideband components is no more than 0.5 milliseconds. A companion platform handles parsing, cleaning, action segmentation and semantic annotation, and links to scene reconstruction, pose estimation and hand tracking. Action retargeting maps human motion to robot bodies. The headband weighs no more than 220 grams and each wristband no more than 75 grams. The tactile glove is optional because some data-collection operators asked that workers not wear gloves, as waterproof tasks such as dishwashing could be affected and sensors could introduce noise; Yang said tests showed pressure readings could persist after gloves touched water cups and were removed. PHANES positions SeePhys Hub in the middle layer of a three-tier data pyramid between large-scale internet video and high-precision motion-capture or real-robot data. The company's earlier work includes EgoTouch, TouchAnything and TouchWorld, which tested plug insertion, pot wiping and tissue pulling. PHANES plans to launch a full-modal world model in 2027, starting with industrial and food-and-retail applications. It does not plan to build its own robot body yet, citing the risk of sunk costs before data, models and deployment scenarios stabilize.

openJiuwen released a personal knowledge space that WorkSwarm now natively supports. The system collects content from sources such as local folders, Edge favorites, Zhihu columns, Toutiao, Feishu documents, GitHub and AtomGit. It supports manual and scheduled incremental collection, unifies formats, identifies new or changed material, and organizes it into topic directories and Markdown pages with document links and provenance. Users can choose rule-based processing using BM25 term weighting and content similarity, a hybrid of rules and large models, or agent-based processing. In agent mode, the system analyzes relationships between new material and existing knowledge, updates or merges existing topic pages, adds documents when necessary, and maintains directory descriptions and cross-references. When a task starts, the agent first receives an overview and reading entry points, then reads relevant pages as needed in a progressive-disclosure manner. Collection and task use can be controlled separately. In a 25-question test set based on local papers, browser bookmarks, Feishu documents, GitHub, AtomGit, Zhihu and Toutiao, openJiuwen reported that enabling the personal knowledge space raised task pass rate by 32%, cut total task time by 76.2% and reduced token consumption during the task phase by 84.1%. Preprocessing also consumes tokens, but the company said lower-cost models can be used for organization and the cost can be amortized across repeated tasks. The feature is available on WorkSwarm.

Editor's Summary

On Oct. 14, SenseTime's Dax platform reported 80% higher dispatch efficiency for Shanghai's first quantum-city agents, PHANES AI launched SeePhys Hub for full-body multimodal robot data collection, and openJiuwen introduced a personal knowledge space for office agents. The announcements show Chinese AI teams connecting large models with city services, robot training data and reusable personal knowledge. They also mark a shift from model demonstrations toward deployable agent and robotics infrastructure.