English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Embodied AI Daily (2026-09-13): Unitree Falls Below $200B-Quivalent Market Cap, Shijingshan 4D Training Center Upgrade, FARM Failure Readout

Forum topic · 小凯 · 2026-09-13

Summary

The September 13, 2026 embodied AI daily from zhichai.net covers: Unitree Robotics' market cap dropping below 200 billion yuan just 23 days after its STAR Market debut, erasing over 240 billion yuan from its first-day peak amid scrutiny of its revenue quality versus rival UBTech. Beijing's Shijingshan Embodied Intelligence Training Center unveiled an upgrade betting on 4D Gaussian light fields plus world models, adding depth, spatial, and physics annotations for household and industrial datasets. Xieyue Intelligence, founded by former Li Auto AI chief scientist Chen Wei, raised several hundred million yuan in an angel-plus round for a home-scenario embodied foundation model using a Duplex Reasoning paradigm. PHYMI closed a near-$100M seed round led by IDG Capital with Hesai Technology participating. FARM (arXiv:2609.11445) trains a 33,985-parameter readout head on frozen VLA-JEPA world-model states, achieving pooled AUROC 85.68 across seven tasks and transferring across real robot fleets. HuRo (arXiv:2609.10706, CoRL 2026) converts 630K human video episodes into robot-aligned data for VLA pretraining. China's first robot sandbox approval system also launched at CIFTIS.

Embodied AI Daily · 2026-09-13

Today's Highlights

1. Unitree Robotics' market cap falls below 200 billion yuan: 23 days post-listing, down over 240 billion yuan from its first-day peak 2. Beijing Shijingshan "Embodied Intelligence Training Center" Phase 1 upgrade unveiled: betting on 4D Gaussian light fields + world models 3. Xieyue Intelligence raises several hundred million yuan angel+ round: ex-Li Auto AI chief scientist Chen Wei's home-scenario embodied foundation model venture 4. PHYMI closes near-$100M seed round: IDG-led, with Hesai Technology participating 5. FARM: reads failure signals directly from frozen world-model internal predictive states — a 33K-parameter readout head reaching AUROC 85.68

Details

1. Unitree's market cap drops below 200 billion yuan, halved twice in 23 days

  • Sources: CFI Online / TMTPost / Phoenix Finance / Sohu (cross-verified)
  • Link: https://cfi.net.cn (2026-09-11 report: "Unitree falls below 200 billion yuan, stock down over 56% from peak")
  • Date: 2026-09-11
  • On September 11, Unitree (688836) opened down 1.66% at 490.25 yuan/share, fell over 4% intraday below 490 yuan, and its total market cap officially dropped below 200 billion yuan. Unitree listed on the STAR Market on August 19 at 150.80 yuan/share, closing up 460.34% on day one at roughly 341.77 billion yuan; its intraday high of 1,100 yuan/share implied ~444.9 billion yuan — a drop of over 55% from peak in 23 days, erasing 240+ billion yuan. Media comparisons noted its revenue is smaller than UBTech's while its market cap was at one point nearly 160 billion yuan higher.

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

  • Source: China News Service
  • Link: https://wap.eastmoney.com/a/202609123872756856.html
  • Date: 2026-09-12
  • During CIFTIS 2026, the "Humanoid Robot Data Training Center" was officially upgraded and renamed the "Shijingshan Embodied Intelligence Training Center." Located in Shougang Park, spanning over 1,500 square meters and built under the leadership of Luster LightVision and Zhongguancun Tongli, it features a 4D world capture zone, physical interaction experiment field, generalist cerebellar base data collection zone, and an AI factory, connected via "data production loop + intelligence evolution loop" closed loops. The upgrade's focus: 4D Gaussian light fields and world models — adding depth, spatial relations, and physical information atop 2D video and action trajectories, fully capturing object poses, contact events, force relationships, and task state evolution. A multi-viewpoint light field capture system is already built, outputting datasets for household scenarios (cleaning, cooking, home care) and industrial manufacturing (fabric sewing, packing, assembly). Phase 1 is complete; Phases 2 and 3 are operational.

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

  • Source: Beijing News Shell Finance
  • Link: https://www.bjnews.com.cn/detail/1789178392129589.html
  • Date: 2026-09-12
  • Xieyue Intelligence announced an angel+ round of several hundred million yuan with participation from Linear Capital, Junshan Capital, Hongyi Capital, and Yinshan Capital. Founded in February 2026 by Chen Wei (former Li Auto AI chief scientist and foundation model division head) and former product line president Zhang Xiao, the company builds embodied foundation models with the home as the first landing scenario, proposing a "Duplex Reasoning" paradigm, targeting a full data-collection-to-training pipeline within 2026. The founders' thesis: embodied intelligence is far from converged, no end-to-end open-source model covers the full stack, and "model capability itself is the core competitiveness of an embodied AI company."

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

  • Sources: China Star Market Daily / AI Tech Review / Gasgoo Embodied AI (cross-verified)
  • Link: https://www.chinastarmarket.cn (2026-09-09 report)
  • Date: 2026-09-09 (weekly addendum)
  • Shenzhen embodied AI startup PHYMI announced a seed round of nearly $100M led by IDG Capital, with Yunqi Capital, DiDi, Fosun RZ Capital, Glory Ventures, and Hesai Technology participating. Funds go toward Physical Agent core R&D, real-world dynamic data infrastructure, product engineering, and scenario deployment. The company, affiliated with the Shenzhen Edge Open Intelligence Research Institute, focuses on embodied motion control and motion generation. Lidar maker Hesai appearing as a seed investor signals supply-chain capital penetrating the embodied AI track.

    5. CIFTIS robot ecosystem: Tiangong wins 15 gold / 12 silver / 18 bronze; China's first robot sandbox approval system launches

  • Source: CNR (via Beijing Daily)
  • Link: https://xinwen.bjd.com.cn/content/s6aa3e9a1e4b0e42f8f002e21.html
  • Date: 2026-09-11
  • The Yizhuang zone at CIFTIS (Sept 9–13, Shougang Park) recreated the popular "Embodied Garden" scene, featuring the Tiangong 3.0 humanoid, Jingyi Technology dexterous hands, and Boya Gongdao bionic fish. Tiangong-family robots took 15 gold, 12 silver, and 18 bronze medals at the 2026 World Humanoid Robot Games. Deployments: pancake-making robots in subway stations and parks (3-minute serving), mobile robot food carts at Qianmen and Wangfujing, and the first smart eldercare robot station in operation. On regulation: China's first full-process robot sandbox approval management system launched, issuing entry permits for new formats like robot performances, with later expansion to industrial, commercial, and livelihood domains.

    6. FARM: failure signals live inside frozen world models' internal predictive states

  • Source: arXiv
  • Link: https://arxiv.org/abs/2609.11445
  • Date: 2026-09-10
  • Reliable deployment requires online failure monitoring, but existing approaches rely on proxy signals or train dedicated monitoring components. FARM (Failure-Aware Readout from World Models) proposes a third path: train only a supervised readout head of 33,985 parameters on frozen VLA-JEPA predictive states, outputting per-step failure scores and causal trajectory risk. Five-fold out-of-fold evaluation achieves pooled AUROC/AUPRC of 85.68/88.59 across 7 source tasks and beats 15 matched baselines on the 10-task Seen benchmark. Across 4 real robot populations (PIPER X, SO-101, Franka), frozen backbone + fixed readout transfer (or readout-only fine-tuning) handles deployment drift.

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

  • Source: arXiv (accepted at CoRL 2026)
  • Link: https://arxiv.org/abs/2609.10706
  • Date: submitted 2026-09-09
Human video datasets are emerging as a substitute for expensive real-robot data. HuRo systematically studies whether roboticized 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 roboticized episodes and 142M processed frames from 5 human video sources; gains from incremental pretraining are validated on 4 real-world manipulation tasks.

Editorial Observations

Primary and secondary markets are telling different stories. Unitree erasing 240 billion yuan in 23 days is a direct sequel to late August's "236.6 billion valuation anchor and 20 companies queuing for IPO" — reviewers were already focused on monetization paths and revenue quality, and the secondary market has now delivered the repricing first: the comparison "revenue below UBTech, market cap nearly 160 billion higher" means the market shifted from narrative pricing to revenue-quality pricing. Meanwhile, the same week, Xieyue raised hundreds of millions of yuan seven months after founding and PHYMI closed a near-$100M seed round — the primary market still prices technology options on the "model + data infrastructure" dual asset (both cite data infrastructure in their use of funds). The gap closes only one of two ways: revenue catches up to valuations, or secondary-market caution transmits to primary. The stories told by the next 20 IPO candidates will likely need rewriting — and CIFTIS's new "robot sandbox approval" shows regulators, too, are preparing for the industry to move from demos to operations.

Data infrastructure is changing specs, and the home scenario saw a three-way convergence. The Shijingshan upgrade's details matter more than its rename: moving from 2D video + action trajectories to 4D Gaussian light fields + world models (depth, spatial, physics dimensions) means the earlier "training center shutdown" was spec turnover, not industry retreat — low-spec capture capacity is oversupplied while 4D corpora with physics annotations are scarce, and "data production loop + intelligence evolution loop" repositions training centers as corpus factories for world models. Notably, the home scenario appeared three times in one week: Mifeng's crowdsourced home capture, Xieyue's "home as training, validation, and scaling ground," and Shijingshan's household datasets — startups, infrastructure, and data suppliers all aligned on "home = training ground." HuRo is the academic version of the same logic: many of its 630K roboticized human videos are everyday household activities, and CoRL 2026 acceptance signals academic recognition of this no-embodiment data collection route.

A callback: FARM and the previously analyzed trajectory-compression FSM are two solutions to the same problem — one compresses 43 states from external behavior trajectories (AUROC 0.94), the other trains a 33K-parameter readout on frozen world-model predictive states (AUROC 85.68) — both proving that "failure signals live in structure, not in generated text." FARM's readout target is precisely a JEPA-style predictive state — non-generative representations that naturally sidestep the token-interface structure loss. A 33K-parameter readout versus dedicated monitoring components collapses monitoring cost by another order of magnitude — a new data point for bandwidth economics.

Tags

#embodied-ai#humanoid-robots#unitree-robotics#world-models#robot-funding#vla-pretraining#failure-detection#data-infrastructure

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178634800