Embodied AI Daily · 2026-09-13

Highlights

1. Unitree Technology's market cap falls below 200 billion CNY: 23 days post-IPO, over 240 billion CNY evaporated from first-day peak 2. Beijing Shijingshan "Embodied Intelligence Training Center" Phase I upgrade debuts: betting on 4D Gaussian light fields + world models 3. Xiejue Intelligence closes several-hundred-million CNY angel+ round: former Li Auto AI chief scientist Chen Wei builds home-scenario embodied foundation models 4. PHYMI raises nearly $100M seed round: IDG-led, with Hesai Technology participating 5. FARM: reads failure signals directly from the internal predictive states of a frozen world model, 33K-parameter readout head achieves 85.68 AUROC

Details

1. Unitree market cap falls below 200 billion CNY, halved twice in 23 days

• Sources: CFI Online / TMTPost / iFeng Finance / Sohu (cross-verified)
• Link: https://cfi.net.cn (2026-09-11 report)
• On the morning of Sep 11, Unitree Technology (688836) opened down 1.66% at 490.25 CNY/share, fell over 4% intraday below 490 CNY, and its total market cap officially dropped below 200 billion CNY. Unitree listed on the STAR Market on Aug 19 at an issue price of 150.80 CNY/share, closing up 460.34% on day one at ~341.77 billion CNY, with an intraday high of 1,100 CNY/share implying ~444.9 billion CNY. In 23 days the stock fell over 55% from peak, erasing over 240 billion CNY. Media comparisons noted its revenue is smaller than UBTech's, yet its market cap was once nearly 160 billion CNY higher.

2. Shijingshan Embodied Intelligence Training Center upgrade: 4D Gaussian light fields + world models

• Source: China News Service — https://wap.eastmoney.com/a/202609123872756856.html (2026-09-12)
• During CIFTIS 2026, the "Humanoid Robot Data Training Center" was upgraded into the "Shijingshan Embodied Intelligence Training Center." Located in Shougang Park, covering 1,500+ square meters and led by Luster LightVision and Zhongguancun Tongli, it includes a 4D world capture zone, physical interaction experimental field, general cerebellum foundation data collection zone, and an AI factory, connected via a "data production loop + intelligence evolution loop." The upgrade emphasizes two technical routes: 4D Gaussian light fields and world models — layering depth, spatial relations, and physical information on top of 2D video and motion trajectories to fully capture object poses, contact events, force relations, and task state evolution. A multi-viewpoint light-field capture system is built, outputting datasets for home scenarios (housework, cooking, family care) and industrial manufacturing (fabric sewing, packing, assembly). Phases II and III are already operational.

3. Xiejue Intelligence raises angel+ round: ex-Li Auto AI chief scientist builds home embodied foundation models

• Source: Beijing News Shell Finance — https://www.bjnews.com.cn/detail/1789178392129589.html (2026-09-12)
• Xiejue Intelligence announced an angel+ round of several hundred million CNY, with Linear Capital, Junshan Capital, Hongyi Capital, and Yinshan Capital participating. Founded in Feb 2026 by Chen Wei (former Li Auto AI chief scientist, foundation model division head) and Zhang Xiao (former product line president), it builds embodied foundation models with home as the first landing scenario, proposing a "Duplex Reasoning" paradigm and aiming to complete the full data collection-cleaning-labeling-training pipeline within 2026. The founders argue embodied intelligence is far from converged — no end-to-end open-source model can dominate the whole stack — and that "model capability itself is the core competitiveness of embodied AI companies."

4. PHYMI closes near-$100M seed round, a recent record

• Sources: STAR Market Daily / AI Tech Review / GaShi (cross-verified) — https://www.chinastarmarket.cn (2026-09-09)
• Shenzhen embodied AI company PHYMI raised a near-$100M seed round led by IDG Capital, with QiYun Capital, DiDi, Fosun RZ Capital, Glory Ventures, and Hesai Technology following. Funds go to Physical Agent core R&D, real dynamic data systems, product engineering, and scenario deployment. Hesai, a lidar vendor, appearing in an embodied AI seed round signals supply-chain capital penetrating the sector.

5. CIFTIS robotics ecosystem: TianGong wins 15 golds, first robot sandbox approval system goes live

• Source: CNR / Beijing Daily — https://xinwen.bjd.com.cn/content/s6aa3e9a1e4b0e42f8f002e21.html (2026-09-11)
• The Yizhuang zone at CIFTIS (Sep 9–13, Shougang Park) replicated the "Embodied Garden" from the World Robot Conference, featuring TianGong 3.0 humanoid robots, Jingshi dexterous hands, and Boyagongdao bionic fish. TianGong robots won 15 gold, 12 silver, and 18 bronze medals at the 2026 World Humanoid Robot Games. Deployments include pancake-making robots at subway stations and parks (3-minute serving), mobile robot food carts at Qianmen and Wangfujing, and the first smart eldercare robot station. On regulation: China's first full-process robot sandbox approval management system launched, issuing access certificates for new formats like robot performances, with later expansion to industrial, commercial, and livelihood domains.

6. FARM: failure signals hide in the internal predictive states of frozen world models

• Source: arXiv — https://arxiv.org/abs/2609.11445 (2026-09-10)
• FARM (Failure-Aware Readout from World Models) trains only a 33,985-parameter supervised readout head on frozen VLA-JEPA predictive states, outputting per-step failure scores and causal trajectory risk. Five-fold out-of-fold evaluation reaches 85.68/88.59 pooled AUROC/AUPRC across 7 source tasks, outperforming 15 matched baselines on a 10-task seen benchmark; across 4 real robot populations (PIPER X, SO-101, Franka), a frozen backbone with fixed or lightly fine-tuned readout handles deployment drift.

7. HuRo: "robotizing" 630K human video episodes for VLA pretraining

• Source: arXiv (accepted to CoRL 2026) — https://arxiv.org/abs/2609.10706 (2026-09-09)
• HuRo systematically studies whether robotized human videos can provide effective, scalable supervision for VLA policy pretraining: its pipeline converts heterogeneous human videos into robot-aligned observations and action trajectories, inferring missing intermediate signals across annotation levels. The resulting HuRo dataset contains ~630K robotized episodes (142M processed frames) from 5 human video sources, with gains validated via incremental pretraining on 4 real manipulation tasks.

Editorial Observations

Primary and secondary markets are telling different stories. Unitree losing 240 billion CNY in 23 days is the direct sequel to the "236.6 billion valuation anchor and 20 IPO candidates" line from late August — the secondary market repriced first, shifting from narrative pricing to revenue-quality pricing. Meanwhile, Xiejue (7 months old) and PHYMI still priced as technology options on the "models + data systems" dual asset in the primary market; both prospectuses explicitly list data infrastructure spending. This temperature gap can only converge via revenue catching up to valuations, or primary caution catching down. The robot sandbox approval system at CIFTIS shows regulators also preparing for the industry to move from demos to operations.

Data infrastructure is changing specs; the home scenario achieved a three-way convergence. Shijingshan's real news is the technology shift from 2D video + trajectories to 4D Gaussian light fields + world models — not industry retreat but spec replacement: low-spec capture capacity is oversupplied while 4D corpora with physical annotation are scarce. The home scenario appeared three times this week: Mifeng's crowdsourced home capture, Xiejue's "home as training, validation, and deployment ground," and Shijingshan's home datasets (housework/cooking/care). HuRo is the academic version of the same logic — the human video internet is becoming a ready-made corpus for VLA pretraining.

A callback: FARM and the previously analyzed trajectory-compression FSM solve the same problem — failure signals live in structure, not in generated text. FARM's readout target (JEPA-style predictive states) naturally avoids token-interface structure loss, and a 33K-parameter head collapses monitoring cost by another order of magnitude.