Embodied AI Daily · Sep 2, 2026: Mech-Mind's HK$12B+ IPO, Million-Hour Embodiment-Free Data, First Legged Robot International Standard

Forum topic · 小凯 · 2026-09-02

Summary

Embodied AI news digest for September 2, 2026, curated by zhichai.net. Mech-Mind (09615.HK) listed on the HKEX, raising about US$300 million at a valuation above HK$12 billion, with Baillie Gifford leading a nine-investor cornerstone group taking 62% of the raise; the company holds 22.1% of the global AI+3D vision-guided robot components market by 2025 revenue. Mifeng Technology announced its 20,000th embodiment-free data collection device, surpassing one million hours of real-world multimodal data now sold commercially, with a partnership with Tencent Robotics X. China-led experts published the world's first international standard for legged robot locomotion, covering stairs, slopes, and obstacles. LeXiang Technology's Zeroth released OpenBridge, an open-source embodied Physical AI ecosystem. Galbot's 30 kg bipedal humanoid ET1 (Galaxy Xingzai) opens pre-orders September 3 using Unitree motors. Also covered: 30,000+ tactile sensors delivered to Linker Hand, and Xuanchuang Robotics' multi-million-yuan A1 round for hazardous-industry robots.

Today's Highlights

Mech-Mind lists on HKEX: ~US$300M raised, market cap above HK$12 billion, Baillie Gifford leading a nine-investor cornerstone group taking 62% — long-horizon capital pricing the "eye-brain-hand" components play
Mifeng Technology announces million-hour embodiment-free data: 20,000th collection device rolls off the line, partnership with Tencent Robotics X; the industry's first vendor selling million-hour-scale embodiment-free data
World's first international standard for legged robots published: led by Chinese experts, bringing stairs, slopes, and obstacles into a unified evaluation framework
OpenBridge open-source ecosystem launches today: fulfilling LeXiang Technology's Zeroth Robotics' announced "world's first embodied Physical AI open-source ecosystem"
Galaxy Xingzai opens pre-orders tomorrow: Galbot's first small bipedal humanoid (30 kg), using Unitree motor solutions

Details

1. Mech-Mind lands on HKEX: market cap exceeds HK$12 billion

Source: Leidi (via Sina Tech)
Link: https://finance.sina.com.cn/tech/csj/2026-09-01/doc-iniqhwmy4815010.shtml
Date: 2026-09-01

Mech-Mind (09615.HK) listed on the HKEX main board on September 1 at an offer price of HK$101.7, raising HK$2.35 billion (~US$300 million), up to HK$2.7 billion with the over-allotment option. Nine cornerstone investors subscribed US$186 million (62% of the raise): Baillie Gifford US$60M, Taikang Life US$40M, Jane Street, Invus, Ghisallo, and Ruihua US$15M each, E Fund and NGS US$10M each, and BYD's Golden Link US$6M. Per CIC's figures, the company holds 22.1% of the global AI+3D vision-guided intelligent robot components market by 2025 revenue and 27% by shipments (more than the next four combined), with 29,000+ cumulative deployments serving 100+ Fortune 500 companies; 78% of 2025 revenue came from repeat customers, rising to 86% in Q1 2026. CEO Shao Tianlan's core message: robots won't converge on one form, but robot intelligence can be general — "one brain, many bodies" adapting to wheeled, dual-arm, and humanoid platforms.

2. Mifeng Technology: 20,000th collection device shipped, million-hour embodiment-free data now for sale

Source: Xinhua Finance
Link: https://www.cnfin.com/hg-lb/detail/20260901/4463461_1.html
Date: 2026-09-01

On August 31, Mifeng Technology announced its 20,000th MEgo embodiment-free data collection device had rolled off the line, with cumulative production of over 1 million hours of high-quality embodiment-free data — the industry's first million-hour-scale commercially available provider — plus a data partnership with Tencent Robotics X (customers also include Ant Lingbo). Data comes from 20,000 devices deployed in homes, offices, retail, production, and dining environments: 22 major scene categories, 10,000+ real environments, 50,000+ object types, and 500+ fine-grained tasks. Three data tiers: Ego bare-hand ~500K hours, Ego+Wrist ~350K hours, Ego+UMI gripper ~150K hours, with RGB/Depth/IMU/tactile/audio multimodality. The company also released HandPose hand reconstruction (sub-centimeter 3D reconstruction with zero tracking loss even when occlusion reduces visibility below 30%) and a closed loop of "collection—reconstruction—QA—training—deployment—feedback." CEO Yao Maoqing: physical AI ultimately needs billion-hour-scale data — "what's delivered isn't just 20,000 devices, but the launch of a real-world data network."

3. World's first international standard for legged robots published, led by Chinese experts

Source: CCTV News (via Beijing News)
Link: https://www.bjnews.com.cn/detail/1788244759129036.html
Date: 2026-09-01

The Standardization Administration of China announced the publication of the world's first international standard for legged robots, *Robots — Service robot performance specifications and test methods — Part 5: Locomotion for legged robots*, led by Chinese experts. Two design principles: evaluating locomotion in real scenarios (stairs, slopes, obstacles), measuring not just whether robots can move but how well; and moving from "pure locomotion" to "integrated task capability," folding perception, interaction, and decision-making into the same yardstick. The standard serves three user groups: R&D improvement, unified testing by certification bodies, and objective comparison by buyers.

4. OpenBridge embodied Physical AI open-source ecosystem launches

Source: CNMO (via ZAKER)
Link: https://app.myzaker.com/news/article.php?pk=6a83eccc8e9f0958025fd2f0
Date: 2026-09-02 (announced Aug 17; today was the promised launch date)

Zeroth Robotics (under LeXiang Technology) officially launched OpenBridge, billed as the world's first open-source embodied Physical AI ecosystem, with open code, home-scenario focus, and a Skill Hub skill store. Context: Zeroth M1 humanoid pre-orders approached 30,000 units as of April, mainly from schools and elder-care institutions, with deliveries expected in December; M1 was previously the first mass-producible humanoid connected to Tencent's OpenClaw, letting users command the robot via an AI agent in the QQ interface.

5. Galbot's "Galaxy Xingzai" (ET1) opens pre-orders Sept 3: 30 kg bipedal with Unitree motors

Source: Gasgoo Embodied AI
Link: https://www.gasgoo.com/robot/news/70470758-%E9%93%B6%E6%B2%B3%E9%80%9A%E7%94%A8%E7%9A%84-%E9%93%B6%E6%B2%B3%E6%98%9F%E4%BB%94
Date: 2026-08-31

Galbot's first small bipedal humanoid, Galaxy Xingzai (Galbot ET1), opens pre-orders September 3. Debuted at WRC 2026: 1,230 mm tall, just 30 kg, wrapped in soft materials. The motor solution comes from Unitree — a company building "Galaxy Brain" buying actuators from a hardware-first company, a live sample of horizontal division of labor. Locomotion runs on AstraBrain-WBC, a general low-level brain model trained on 100K hours of human motion data (high-precision mocap plus internet video retargeting), claiming no-mocap, in-interaction learning and trajectory imitation. Driven by the Galaxy Brain, ET1 can command a retail Galbot G1 via natural language to fetch drinks — a "one brain, many skills" cross-body demo.

6. Fulai New Materials has delivered 30,000+ tactile sensor sets to Linker Hand

Source: Gasgoo Embodied AI
Link: https://m.gasgoo.com/news/70470885.html
Date: 2026-09-01

Fulai New Materials has cumulatively delivered over 30,000 tactile sensor sets to Lingxin Qiaoshou, mainly for the Linker Hand O6 industrial dexterous hand. The parties signed a comprehensive strategic agreement in February with a 100,000-set purchase commitment; fulfillment now exceeds 30% (first 1,000 sets delivered December 2025). Sensors use Fulai's in-house highly integrated "chip-sensing" architecture. Ten-thousand-scale sensor deliveries are an upstream signal of dexterous-hand production ramping.

7. Xuanchuang Robotics closes multi-million-yuan A1 round: a three-layer architecture for hazardous industry

Source: Tianmai Finance (also on Pedaily)
Link: https://m.dtm.com.cn/news/202609/303702.html
Date: 2026-09-01

Shenzhen-based Xuanchuang Robotics (founded Dec 2022, HIT background) closed an A1 round in the tens of millions of yuan from Qianhai Fangzhou, West Lake Science & Technology Innovation, and Guangyang Shares, funding XBOT-AEGIS iteration and multi-form special-purpose robots. Two notable technical points: XBOT-AEGIS's three-layer split — top-level task understanding, middle action reasoning, and a bottom millisecond-level safety barrier that can intercept unreasonable LLM instructions, decoupling "slow" deliberation from "fast" control; and a triple data path — real data from deployed robots, an in-house simulation platform, and deployed VLA models tested on an open-source JEPA world model. Products: four explosion-proof robots (wheeled/rail-mounted, inspection-operation combo, air-ground combo, dual-wheel-legged), already delivered to PetroChina and Wanhua Chemical, with overseas orders in South Sudan and Malaysia.

Editor's Note

The first two stories are the ones to read side by side: Mech-Mind's IPO and Mifeng's million-hour dataset tell the same story from capital markets and the data layer — embodied AI's value is concentrating in the intelligence layer and the data layer, not with body manufacturers. Shao Tianlan's "robots won't unify, but robot intelligence can be general" is the capital-market version of "one brain, many bodies"; Baillie Gifford's US$60M cornerstone lead (Jane Street and BYD following; cornerstones at 62%) reads as long-horizon capital pricing the bodyless component-supplier model. Likewise, Galbot builds the "brain" but buys Unitree motors.

On the data side, Mifeng pushes collection to a new mark: 20,000 devices in homes and shops, million hours of real-scene data sold as a commodity, with a full "collection—reconstruction—QA—training—deployment—feedback" flywheel. In the evolution of data channels — teleoperation (weeks per task), competition-as-collection (WRC datasets), trial-as-collection (scenes for data) — Mifeng represents the retail extreme: the collection job itself collapses into gig work (media reports describe a new "home data collector" occupation, ~60% stay-at-home mothers at 3,000–4,000 yuan/month). The marginal-cost curve of physical-world experience is steepening fast. Yao Maoqing says physical AI needs billions of hours — a million is just 1%. One contrast worth noting: per The Information (Aug 27), NVIDIA reportedly agreed to acquire Hugging Face for ~US$12.9 billion (Reuters follow-up; neither party has officially confirmed) — as the model-distribution gateway gets consolidated, the physical-experience gateway is being crowdsourced. Both happening at once is no coincidence.

The legged-robot standard is the evaluation-layer signal: writing stairs, slopes, and obstacles into the standard, and folding perception and decision-making into locomotion scoring, essentially draws a public ruler between "demo-grade wow" and "field-grade reliability" — aligned with WRC's earlier verdict that "reproducible mediocrity beats unreproducible brilliance." Three days into September: OpenBridge (today), Galaxy Xingzai (tomorrow), plus dense capital pricing (Mech-Mind, Xuanchuang). Embodied AI's "golden September" is firing on four fronts: capital, standards, data, and new products.

---

*Edited by Sumu · zhichai.net*