The next USB/MCP for egocentric

If we do not cooperate and work together, we cannot fight against these giants who are already so much ahead of the game.

Research and citation checks for this piece were done by Claude Fable 5.1; every source was opened and verified on 3 September 2026, and the text was reviewed by mutual.

This is the briefing behind a roundtable we are hosting this week with about ten small builders of egocentric capture hardware for robot learning. The short version, in our view: two kinds of giants are far ahead of all of us, none of us is the other’s problem, and the one lever a small builder actually controls is whether its device can record next to someone else’s.

The room and the giants #

The giants come in two kinds, and the biggest of them has both China’s scale and America’s money. The first kind is the well-funded robotics-data company building its own capture stack. Mecka AI is not simply an American company: it was founded in Toronto by a Chinese-Canadian CEO, Josh Gao, an Ivey graduate whose parents were born in China, with a mostly Canadian engineering team [75]; it is headquartered in New York and has raised $60 million across a $25 million Series A closed in November 2025 and a $35 million follow-on [1]; and its custom capture hardware is made in a factory in Shenzhen [1]. XDOF left stealth on 17 June 2026 with $70 million from Thrive Capital, Spark Capital, a16z, Lux and WndrCo [2]. The second is the Chinese player with scale, state backing and a much lower price. 51WORLD put its AperEgo head-mounted kit on sale on 18 August 2026 at an introductory RMB 5,100 per set [3], about $760 at the Federal Reserve’s 28 August rate of 6.7260 yuan per dollar [74]. Mifeng, an AgiBot subsidiary [22] founded in February 2026, says its 20,000th MEgo capture device came off the line on 31 August [4]. We think nobody in the room is competing with the person on the next tile; all of us are competing with these four and the world behind them.

Mecka AI #

Mecka is the clearest case of China’s scale joined to America’s money. It launched from stealth in August 2025 as a Toronto-based Canadian startup, led by Gao, a Canadian whose parents were born in China, with a team of mostly Canadian engineers [75]. By mid-2026 it was headquartered in New York [1], had opened a New York office to grow a customer base that is mostly American, with more than half of its 45 people still Canadian [9], and was designing and building its hardware in China. Business Insider puts Mecka’s total funding at $68 million per the company and reports it recently crossed $100 million in annualized revenue [5]. When Fortune reached CEO Josh Gao he was in Shenzhen, visiting a factory producing custom capture equipment Mecka designed [1]. Investor Kindred describes Mecka as designing and manufacturing its own capture hardware, body sensors and tactile gloves while also using commodity iPhones [6]. The job board shows the build: a Shenzhen calibration engineer owns intrinsics, extrinsics, IMU calibration and multi-sensor alignment for stereo and multi-camera rigs across Mecka’s fleet [7], and a Shenzhen tactile-gloves lead takes a glove from concept through EVT, DVT and PVT to volume with China-based suppliers [8]. Mecka says it sends iPhones and custom cameras to hundreds of thousands of contributors across 12 countries [9]. In the EgoVerse paper, where Gao is a listed author, industry-partner rigs are head-mounted stereo fisheye RGB cameras with synchronized sensing modules, and one partner’s rig is a 6 cm baseline pair recording synchronized 1920 by 1200 at 30 FPS with a tightly time-aligned IMU [10]. Kindred states the requirement plainly: visual and haptic channels must be time-aligned to millisecond precision or a policy learns the wrong thing [6].

XDOF #

XDOF was launched in October 2024 by Philipp Wu, Fred Shentu and Nemo Jin, and Wu and Shentu built GELLO together at Berkeley [2], a teleoperation device with under $300 of parts [11]. By June 2026 it had about 60 employees and 20 customers including unnamed frontier labs [2]. Its data pyramid is teleop on the deployed robot, generalised teleop, and egocentric human data for which XDOF plans to build its own wearable sensors [2], proprietary so that captured data matches its own hand-tracking algorithms [12]. Its open ABC-130K dataset, 3,500 hours and over 130K episodes across 195 tasks [13], was collected on an $8,000 bimanual station that the paper contrasts with the $30,000 AgiBot G1 [14]. Its Perception role requires temporal alignment across cameras, IMU, data gloves and force sensors, and it lists proprietary camera arrays and data gloves among its platforms [15]. After more successful demonstrations made a T-shirt-folding policy worse, XDOF concluded the next challenge is scaling data quality, not volume [16].

China’s kits and their prices #

A China Galaxy Securities note from WAIC 2026 said tactile gloves, UMI grippers and headset-plus-phone ego kits had multiplied, some vendors held real orders of several thousand units, prices were falling, and building a UMI or ego kit is not hard [17].

RMB figures are as published; the dollar figures beside them use the Federal Reserve H.10 rate of 6.7260 yuan per dollar for 28 August 2026 [74], rounded. X Square Robot, valued above RMB 20 billion, about $3.0 billion, after a Series C [26], claims its QUANXTA Zero-G1 headband syncs sensors within 1 ms with 100% frame-level alignment [27]. JD plans to mobilise up to 600,000 people to gather 10 million hours of real-scene video in two years [28]. Mifeng reports 1 million hours of body-less data and a supply deal with Tencent Robotics X [4].

China’s scale and state backing #

CAICT counts more than 70 embodied-AI training grounds operating nationwide and 40-odd more under construction or planned [29]. The Economist, via The Business Standard, reports the government covers about 80% of training-centre cost and that physical data collection runs 500 to 700 yuan, about $74 to $104, per hour [30]. Reuters found a Liuzhou centre supplied under an $18 million regional tender whose staff said it has no clear path to profitability, citing high operating costs and low data prices [31]. Beijing E-Town issues 100 million yuan, about $14.9 million, of data vouchers a year [32], and Shanghai offers up to 5 million yuan, about $740,000, per year in corpus vouchers [33]. MIIT and SASAC target a 10,000-unit deployment capability by end-2026 [34]. Interact Analysis puts Chinese vendors above 90% of 2025 humanoid production and names government material support as the primary reason [35]. MIIT approved YD/T 6771-2026 on embodied dataset quality, effective 1 November 2026, and the same report calls one million hours of data capacity a hard threshold for model developers [36]. Teleop data that cost $340 per hour in 2024 now costs $118, and Chinese manufacturers make eight of the fourteen sub-$10K arms on the market [37].

Where “affordable” already sits #

Our read is that the parts floor for a synchronized stereo head unit is under $200 and the shelf price of a finished one is converging on $500. Ego-OSCAR, an open head unit with a hardware-synchronised global-shutter stereo pair and IMU, has a complete bill of materials under USD 200 [38].

The AoE authors’ cost bands agree: UMI-style rigs $300 to $800, wearables over $2k, teleoperation over $50k [48]. If price were our differentiation we would already have lost, and we think that is true of every device in the room.

What none of them has solved #

Our read is that none of these stacks lets a device from one vendor record on the same clock as a device from another. XDOF’s dataset card says ABC-130K streams are sampled on independent clocks and users must nearest-neighbour match timestamps [49], and its recording code holds tick rates with Python sleep and a 300 microsecond busy-spin [14]. The original UMI syncs two GoPros by scanning a rolling QR code, to plus or minus one sixtieth of a second [42]. EgoScale’s spec lists time-stamped streams and no hardware trigger, PTP or cross-device sync [44]. Each Chinese kit claims tight sync, and each claim, as far as we can read it, stays inside one vendor’s box: 1 ms for X Square [27], a 500 Hz IMU hardware clock for KanKan [24], sub-millisecond wireless for Mifeng [22], 5 ms multi-device for Xense’s tactile gripper [50]. Aria Gen 2 broadcasts timing by sub-GHz radio at sub-millisecond accuracy, but to other Aria Gen 2 or compatible devices [51]. The ISO drafters say humanoid data arrives with differing sensor configurations, timing conventions and coordinate frames [52], and ISO/CD 26264-1 is a dataset life-cycle framework, not a capture or sync spec [53]. EgoKit’s authors describe commercial ego rigs as each tied to a proprietary stack that accepts no other host [54].

The standards already exist: IEEE 1588 PTP supports sub-microsecond synchronisation [55], and IEEE 802.1AS gPTP is a 1588 profile for audio, video and time-sensitive control over 802 LANs [56]. gPTP runs over Wi-Fi through 802.11 Fine Timing Measurement, and an Avnu testbed measured 0.62 µs average offset across an Ethernet-Wi-Fi-Ethernet chain [57]. On a cable, a 3.3 V or 5 V TTL trigger starts exposures within microseconds [58], Luxonis FSYNC typically holds under 10 µs across camera sensors and other triggered hardware [59], RealSense syncs over two wires with a 100 µs 1.8 V pulse [60], and ZED X Nano hardware-syncs over GMSL2 [43]. For the file, MCAP stores schemas beside data and is the default log format in ROS 2 [61]. We think the pieces are on the shelf, and what is missing is an agreement between two small vendors to use them the same way. For a 42-minute walkthrough of when time synchronization is and is not needed, where clock error comes from, and how gPTP works, we point everyone in the room to Intrepid Control Systems’ five-minute gPTP explainer [77] and, for the full picture, Colt Correa’s 42-minute Intrepid webinar and its slides [76].

Precedents for small players who cooperated #

USB was developed by Intel with Compaq, DEC, IBM, Microsoft, NEC and Northern Telecom, who formed the non-profit USB-IF in 1995; a zero-royalty adopters agreement and free compliance workshops took it from zero to about 160 companies in a month and over 800 by 2014 [62]. Five companies formed the Bluetooth SIG on 20 May 1998 [63]; more than 43,000 have since joined [64]. Thirty-four companies announced the Open Handset Alliance on 5 November 2007 [65]; Android held 67.61% of mobile OS usage in August 2026 [66]. Forty companies came to the first RISC-V workshop in January 2015, and a billion RISC-V cores shipped in 2024 alone [67]. ROS began when the leaders of about a dozen open-source robotics frameworks agreed in a two-day workshop to start one cleanly licensed codebase [68]. In our own field, Open X-Embodiment pooled 60 datasets from 34 labs into one format [69], DROID ran one hardware stack across 13 institutions [70], and EgoVerse exists because human datasets were fragmented across institutions [10]. We think cooperation is the normal way an interface wins, not the exception.

The most recent precedent is less than two years old. Anthropic open-sourced the Model Context Protocol on 25 November 2024 as a universal, open standard for connecting AI systems with data sources, replacing fragmented integrations with a single protocol [71]. Four months later OpenAI said it would add MCP support across its products, its CEO writing “People love MCP” [72], and two weeks after that the CEO of Google DeepMind said MCP “is rapidly becoming an open standard for the AI agentic era” and that Google would support it in Gemini [73]. We think that is the shape to copy: an interface published openly first and adopted by rivals within months because it was open and already worked.

What we propose #

Three agenda items, each small enough to finish before the next call.

1. A shared clock convention. Every device accepts an external time reference, reports its timestamps in that reference, and states its sync error. We prefer a gPTP-derived time base because the standard already covers Ethernet and Wi-Fi [56], with a TTL trigger line as the wired fallback [58].
2. Interoperability across vendors. One container and one stream-naming rule, so a glove, a stereo head unit and a UMI gripper from three companies land in one file. We default to MCAP because ROS 2 already writes it [61].
3. A minimum two-vendor interop test. Two devices, two vendors, one session, one file, and a published residual time offset. We think one passed test is worth more than any spec document.

We are not asking anyone to give up a product. We are asking for the boring layer to be common so the interesting layers can compete. USB did that for ports and MCP did it for tools. Egocentric capture needs its own.

A note on sources #

The research behind this piece was done by Claude Fable 5.1 (Anthropic) on 3 September 2026: six research lanes, each one re-checked by a separate fact-checking agent that opened every cited URL and tried to refute the claim, then one synthesis and one final citation-by-citation audit, fourteen agents in all. Claims that failed verification were dropped, and the seven corrections from the final audit were applied before publication. mutual reviewed the text and owns every opinion in it. Every figure above is tied to a source accessed on 3 September 2026. Chinese-language sources are cited by original title with an English gloss. RMB prices are given as published; the dollar figures beside them use the Federal Reserve’s H.10 weekly release of 31 August 2026, which lists 6.7260 yuan per U.S. dollar for 28 August 2026 [74], and are rounded to two or three significant figures. Sync precision, unit counts and revenue figures taken from press releases, company sites and job postings are reported as the company’s own statement. Sentences that start “we think” or “our read” are opinion.

Sources #

[1] Fortune (Ben Weiss), Exclusive: Mecka AI raises $60 million to train robots with human data sourced from body sensors and iPhones, 2026-06-01. https://fortune.com/2026/06/01/mecka-ai-series-a-60-million-robotics-data-training/ (accessed 2026-09-03)

[2] TechCrunch (Tim Fernholz), Collecting robot training data is dirty, unglamorous work. Some AI labs are already paying XDOF to do it., 2026-06-17. https://techcrunch.com/2026/06/17/collecting-robot-training-data-is-dirty-unglamorous-work-some-ai-labs-are-already-paying-xdof-to-do-it/ (accessed 2026-09-03)

[3] PR Newswire (51WORLD company release), 51WORLD Unveils Two Embodied AI Products and Reveals Aerospace Strategy from Low Altitude to Deep Space, 2026-08-19. https://www.prnewswire.com/news-releases/51world-unveils-two-embodied-ai-products-and-reveals-aerospace-strategy-from-low-altitude-to-deep-space-302855142.html (accessed 2026-09-03)

[4] 亚洲资本网 (reprint of 上观新闻), 觅蜂科技交付第20000套MEgo，物理AI数据"电网"立起第一根电线杆 (Mifeng delivers its 20,000th MEgo; the first pole of the physical-AI data “grid” goes up), 2026-09-01. http://zhonggcapitalwangw.yzbytv.com/zhonggcapitalwangw/20260901/0901371896.html (accessed 2026-09-03)

[5] Business Insider (Rya Jetha), 25 Robotics Startups to Watch in 2026, According to Investors, 2026-08-17. https://www.businessinsider.com/robotics-tech-ai-startups-investors-funding-2026-8 (accessed 2026-09-03)

[6] Kindred Ventures (Steve Jang, Yash Kishore), Mecka: Creating the Data Corpus for Physical AI, 2026-06-01. https://kindredventures.com/announcement/mecka-creating-the-data-corpus-for-physical-ai/ (accessed 2026-09-03)

[7] Mecka AI (Ashby job board, posting API), Calibration Engineer, Camera & Sensors, 2026-09-01. https://api.ashbyhq.com/posting-api/job-board/mecka.ai (accessed 2026-09-03)

[8] Mecka AI (Ashby job board), Tactile Gloves Program Lead @ Mecka, 2026-06-01. https://jobs.ashbyhq.com/mecka.ai/d2c9e10f-4b56-4204-908c-b14cde6852b5 (accessed 2026-09-03)

[9] BetaKit (Madison McLauchlan), Mecka AI acquires Docula as it builds the data layer for robotics, 2026-06-29. https://betakit.com/mecka-ai-acquires-docula-as-it-builds-the-data-layer-for-robotics/ (accessed 2026-09-03)

[10] arXiv (2604.07607v2), EgoVerse: An Egocentric Human Dataset for Robot Learning from Around the World, 2026-04-08 (v2 2026-07-07). https://arxiv.org/html/2604.07607v2 (accessed 2026-09-03)

[11] Philipp Wu (project site), GELLO: A General, Low-Cost, and Intuitive Teleoperation Framework for Robot Manipulators, undated. https://wuphilipp.github.io/gello_site/ (accessed 2026-09-03)

[12] SiliconANGLE (Mike Wheatley), Robotic teleoperation data startup XDOF launches with $70M in funding, 2026-06-17. https://siliconangle.com/2026/06/17/robotic-teleoperation-data-startup-xdof-launches-70m-funding/ (accessed 2026-09-03)

[13] arXiv (2606.27375), Scalable Behavior Cloning with Open Data, Training, and Evaluation, 2026-06-25. https://arxiv.org/abs/2606.27375 (accessed 2026-09-03)

[14] arXiv (2606.27375, HTML full text), Scalable Behavior Cloning with Open Data, Training, and Evaluation, undated. https://arxiv.org/html/2606.27375 (accessed 2026-09-03)

[15] XDOF careers, Member of Technical Staff, Perception, undated. https://www.xdof.ai/careers/member-of-technical-staff-perception-e3c4b9ed-94ca-40b1-a432-d8eda69949af (accessed 2026-09-03)

[16] XDOF blog, Lowering the Noise Floor on Robot Data, 2026-06-30. https://www.xdof.ai/blog/warp-rm (accessed 2026-09-03)

[17] 智通财经 via Futu News (relaying China Galaxy Securities), 中国银河证券：数采方案供应增加 具身企业商业化应用竞逐升级 (China Galaxy Securities: supply of data-collection solutions rises as embodied-AI firms race to commercialise), 2026-07-22. https://news.futunn.com/post/76393716/china-galaxy-securities-increased-supply-of-data-acquisition-solutions-intensifies (accessed 2026-09-03)

[18] 51WORLD (company site), 51WORLD | AperData (AperData product page), undated. https://www.51world.com.cn/aperdata (accessed 2026-09-03)

[19] 智东西 / 机器人前瞻 (刘俐杉), 51WORLD发布全新数据底座+应用平台，数据采集效率翻10倍 (51WORLD launches a new data base and application platform; collection efficiency up 10x), 2026-08-18. https://www.zhidx.com/p/585910.html (accessed 2026-09-03)

[20] 电子发烧友 (reprint of 穹明智能), 穹明智能正式开源面向双臂机器人的毫米级高精度数据采集手持系统UMI ver.2 (Qiongming open-sources UMI ver.2, a millimetre-precision handheld collection system for dual-arm robots), 2026-07-04. https://m.elecfans.com/article/8052202.html (accessed 2026-09-03)

[21] Noitom (company shop), Hi5 2.0 PC - VR 版 (Hi5 2.0 PC-VR edition), undated. https://shop.noitom.com.cn/html/mobile/258.html (accessed 2026-09-03)

[22] 36氪 (王毓婵), 智元旗下觅蜂发布一站式物理 AI 数据服务平台｜最前线 (AgiBot subsidiary Mifeng launches a one-stop physical-AI data service platform), 2026-04-16. https://www.36kr.com/p/3769501816439555 (accessed 2026-09-03)

[23] IT之家 (sponsored content), 觅蜂MEgo无本体数采：如何帮企业省下70%数据成本 (Mifeng MEgo body-less capture: how it saves firms 70% of data cost), 2026-08-07. https://www.ithome.com/0/987/027.htm (accessed 2026-09-03)

[24] 砍柴网 (reprint of company release), 重磅发布｜瞰瞰智能 G1 数采眼镜，以车规级精度开启物理 AI 全民数据共创新时代 (KanKan Smart G1 data-collection glasses: automotive-grade precision for mass physical-AI data co-creation), 2026-08-22. https://news.ikanchai.com/2026/0822/665340.shtml (accessed 2026-09-03)

[25] JD Cloud (company site), JoyEgoCam - 具身智能数据交易平台 (JoyEgoCam, Embodied AI Data Exchange Platform), undated. https://robotdata.jdcloud.com/data-collection (accessed 2026-09-03)

[26] Robotics & Automation News (Sam Francis), X Square Robot builds a full-stack approach to embodied AI and general-purpose robotics, 2026-07-03. https://roboticsandautomationnews.com/2026/07/03/x-square-robot-builds-a-full-stack-approach-to-embodied-ai-and-general-purpose-robotics/102954/ (accessed 2026-09-03)

[27] X Square Robot (company site), X Square Robot Launches QUANXTA Zero Series, Reframing Embodied Data Production Through Universal Manipulation Interfaces, 2026-07-02. https://x2robot.com/en/news/6a46341cc7feadddbc603a33 (accessed 2026-09-03)

[28] 央广网 (labelled commercial information), 京东发布行业首个具身数据全链路基础设施 (JD releases the industry’s first full-chain embodied-data infrastructure), 2026-04-17. https://tech.cnr.cn/techph/20260417/t20260417_527589204.shtml (accessed 2026-09-03)

[29] 科技日报 (relaying 央视新闻), 超70家训练场建成启用 具身智能加速产业落地 (Over 70 training grounds built and in use; embodied AI accelerates industrial deployment), 2026-08-19. https://www.stdaily.com/web/gdxw/2026-08/19/content_565996.html (accessed 2026-09-03)

[30] The Business Standard (syndicating The Economist), China is training thousands of robots for the real world, 2026-08-30. https://www.tbsnews.net/worldbiz/china/china-training-thousands-robots-real-world-1528301 (accessed 2026-09-03)

[31] Reuters (Laurie Chen), China can build kung fu-fighting robots. But it can’t get them to do factory work, 2026-08-27. https://www.reuters.com/investigations/chinas-humanoid-robots-arent-smart-enough-take-your-job-yet-2026-08-27/ (accessed 2026-09-03)

[32] 北京经济技术开发区管理委员会, 北京经济技术开发区关于推动具身智能机器人创新发展的若干措施 (Beijing E-Town measures to promote innovative development of embodied-AI robots), 2025-08-12. https://kfqgw.beijing.gov.cn/zwgkkfq/2024zcwj/202508/t20250812_4171688.html (accessed 2026-09-03)

[33] 上海市人民政府, 上海市人民政府办公厅关于印发《上海市具身智能产业发展实施方案》的通知 (Notice issuing the Shanghai Embodied-AI Industry Development Implementation Plan), 2025-08-06. https://www.shanghai.gov.cn/nw12344/20250806/f9cb53544505426d807055ca20bd69fc.html (accessed 2026-09-03)

[34] 工业和信息化部 / 国务院国资委, 工业和信息化部办公厅 国务院国资委办公厅关于联合开展2026年度人形机器人与具身智能实景实训专项行动的通知 (MIIT and SASAC notice on the 2026 humanoid robot and embodied-AI real-scene training special action), 2026-06-08. https://www.miit.gov.cn/jgsj/kjs/wjfb/art/2026/art_cd666691abf8471fb8553d463aa416e3.html (accessed 2026-09-03)

[35] Interact Analysis (Marco Wang), Humanoid robot production surges tenfold in 2025, but commercial deployments remain limited, 2026-06-11. https://interactanalysis.com/insight/humanoid-robot-production-surges/ (accessed 2026-09-03)

[36] 人民邮电报 via 数字中国建设峰会 (szzg.gov.cn), 首个具身智能数据集质量标准发布！ (First embodied-AI dataset quality standard released), 2026-08-04. https://www.szzg.gov.cn/2026/xwzx/szkx/202608/t20260804_5354035.htm (accessed 2026-09-03)

[37] SVRC Research (mirrored on Taiwan STPI outlook portal), State of Robotics 2026, Silicon Valley Robotics Center Annual Report v1.0, 2026-03. https://outlook.stpi.niar.org.tw/pdfview/tdop/2c94808a9e679eca019e865f973b6503 (accessed 2026-09-03)

[38] arXiv (2608.08285, FPV Labs), Ego-OSCAR: Egocentric Open source Stereo CAptuRe System, 2026-08-08 (v2 2026-08-12). https://arxiv.org/html/2608.08285 (accessed 2026-09-03)

[39] DHgate marketplace listing (ELP brand), ELP 3D Stereo USB Camera Module with Global Shutter Dual M12 No Distortion Lenses 3200×1200@60fps, undated. https://www.dhgate.com/goods/1072511160.html (accessed 2026-09-03)

[40] Orbbec (official store), Orbbec Gemini 335 | Active & Passive Stereo 3D Camera for AMRs, undated. https://store.orbbec.com/products/gemini-335 (accessed 2026-09-03)

[41] RealSense (official store), RealSense Depth Camera D435i, undated. https://store.realsenseai.com/buy-intel-realsense-depth-camera-d435i.html (accessed 2026-09-03)

[42] arXiv (2402.10329, Chi et al.), Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots, undated. https://arxiv.org/html/2402.10329 (accessed 2026-09-03)

[43] Stereolabs (store), ZED X Nano Stereo Camera, undated. https://www.stereolabs.com/en-gb/store/products/zed-x-nano-stereo-camera (accessed 2026-09-03)

[44] EgoScale, Capture C1 | Stereo Egocentric Capture Hardware, undated. https://www.egoscale.com/capture-c1 (accessed 2026-09-03)

[45] General Intelligence Labs, EGO1: An All-in-One Headset for Egocentric Robot-Learning Data, 2026-06-02. https://www.gilabs.xyz/blog/ego1 (accessed 2026-09-03)

[46] AIFITLAB (reseller), General Intelligence Labs Ego1, undated. https://aifitlab.com/products/general-intelligence-labs-ego1 (accessed 2026-09-03)

[47] Meta / Project Aria, Project Aria Research Kit, undated. https://www.projectaria.com/research-kit/ (accessed 2026-09-03)

[48] arXiv (2602.23893, Yang et al.), AoE: Always-on Egocentric Human Video Collection for Embodied AI, 2026-02-27. https://arxiv.org/html/2602.23893 (accessed 2026-09-03)

[49] Hugging Face (XDOF org), XDOF/ABC-130k, Datasets at Hugging Face, undated. https://huggingface.co/datasets/XDOF/ABC-130k (accessed 2026-09-03)

[50] 千觉机器人 Xense Robotics (company site), XTAC UMI G1-穿戴式视触觉多模态数据采集夹爪 (XTac UMI G1 wearable visuotactile multimodal data-collection gripper), undated. https://xenserobotics.com/product/392224/detail/20 (accessed 2026-09-03)

[51] Meta AI blog, Inside Aria Gen 2: Explore the Cutting-Edge Tech Behind the Device, 2025-06-04. https://ai.meta.com/blog/aria-gen-2-research-glasses-under-the-hood-reality-labs/ (accessed 2026-09-03)

[52] arXiv (2606.19769, Shaoshan Liu et al., ISO/TC 299/WG 16 contributors), Data Standards for Humanoid Robotics: The Missing Infrastructure for Physical AI, 2026-06-18. https://arxiv.org/html/2606.19769 (accessed 2026-09-03)

[53] Institute for Standardization of Serbia (mirror of ISO project record), ISO/CD 26264-1 Humanoid robot datasets, Part 1: General requirements, stage 30.60 effective 2026-08-26. https://iss.rs/en/project/show/iso:proj:93011 (accessed 2026-09-03)

[54] arXiv (2605.16797v1, George Mason University, ByteDance et al.), EgoKit: Towards Unified Low-Cost Egocentric Data Collection with Heterogeneous Devices, 2026-05. https://arxiv.org/html/2605.16797v1 (accessed 2026-09-03)

[55] IEEE Standards Association, IEEE 1588-2019, IEEE Standard for a Precision Clock Synchronization Protocol for Networked Measurement and Control Systems, undated. https://standards.ieee.org/ieee/1588/6825/ (accessed 2026-09-03)

[56] IEEE Standards Association, IEEE 802.1AS-2020, Timing and Synchronization for Time-Sensitive Applications, undated. https://standards.ieee.org/ieee/802.1AS/7121/ (accessed 2026-09-03)

[57] Avnu Alliance (Fatemeh Fazel, Dave Cavalcanti, eds.), Wireless TSN: Tests and Use Case Demonstrations with Wi-Fi, Version 1.0, 2024-01. https://avnu.org/wp-content/uploads/2024/02/Avnu-WTSN-Tests-and-Use-Case-Demonstrations-with-Wi-FI-white-paper-v1.0-approved.pdf (accessed 2026-09-03)

[58] Teledyne Vision Solutions (FLIR), Configuring Synchronized Capture with Multiple Cameras, undated. https://www.teledynevisionsolutions.com/support/support-center/application-note/iis/configuring-synchronized-capture-with-multiple-cameras/ (accessed 2026-09-03)

[59] Luxonis documentation, Hardware synchronization (FSYNC), undated. https://docs.luxonis.com/hardware/platform/deploy/data-sync/hardware-sync.md (accessed 2026-09-03)

[60] RealSense developer documentation, Multi-Camera configurations, D400 Series Stereo Cameras, 2026-02-04. https://dev.realsenseai.com/docs/multiple-depth-cameras-configuration/ (accessed 2026-09-03)

[61] Foxglove / MCAP project, MCAP, undated. https://mcap.dev/ (accessed 2026-09-03)

[62] Intel Corporation (Jeannine Drew), Two decades of “plug and play”: How USB became the most successful interface in the history of computing, undated. https://www.intel.com/content/dam/www/public/us/en/documents/articles/usb-two-decades-of-plug-and-play-article.pdf (accessed 2026-09-03)

[63] Bluetooth SIG, 20 years of blue: Experience the interactive history of Bluetooth, undated. https://www.bluetooth.com/wp-content/uploads/Sitecore-Media-Library/20year/default.html (accessed 2026-09-03)

[64] Bluetooth SIG (Jason Marcel), Understanding membership at the Bluetooth SIG, 2026-04-29. https://www.bluetooth.com/blog/understanding-membership-at-the-bluetooth-sig/ (accessed 2026-09-03)

[65] Open Handset Alliance (press release), Industry Leaders Announce Open Platform for Mobile Devices, 2007-11-05. https://www.openhandsetalliance.com/press_110507.html (accessed 2026-09-03)

[66] Statcounter Global Stats, Mobile Operating System Market Share Worldwide, August 2026. https://gs.statcounter.com/os-market-share/mobile/worldwide (accessed 2026-09-03)

[67] RISC-V International, RISC-V Annual Report 2025, 2026-01. https://riscv.org/wp-content/uploads/2026/01/RISC-V-Annual-Report-2025.pdf (accessed 2026-09-03)

[68] IEEE Spectrum (Keenan Wyrobek), The Origin Story of ROS, the Linux of Robotics, 2017-10-31. https://spectrum.ieee.org/the-origin-story-of-ros-the-linux-of-robotics (accessed 2026-09-03)

[69] arXiv (2310.08864, Open X-Embodiment Collaboration), Open X-Embodiment: Robotic Learning Datasets and RT-X Models, 2023-10-13. https://arxiv.org/html/2310.08864 (accessed 2026-09-03)

[70] DROID Dataset Team (project site), DROID: A Large-Scale In-the-Wild Robot Manipulation Dataset, undated. https://droid-dataset.github.io/ (accessed 2026-09-03)

[71] Anthropic, Introducing the Model Context Protocol, 2024-11-25. https://www.anthropic.com/news/model-context-protocol (accessed 2026-09-03)

[72] TechCrunch (Kyle Wiggers), OpenAI adopts rival Anthropic’s standard for connecting AI models to data, 2025-03-26. https://techcrunch.com/2025/03/26/openai-adopts-rival-anthropics-standard-for-connecting-ai-models-to-data/ (accessed 2026-09-03)

[73] TechCrunch (Kyle Wiggers), Google to embrace Anthropic’s standard for connecting AI models to data, 2025-04-09. https://techcrunch.com/2025/04/09/google-says-itll-embrace-anthropics-standard-for-connecting-ai-models-to-data/ (accessed 2026-09-03)

[74] Board of Governors of the Federal Reserve System, Foreign Exchange Rates, H.10 Weekly, release date 2026-08-31 (China, P.R., yuan: 6.7260 per U.S. dollar on 2026-08-28). https://www.federalreserve.gov/releases/h10/current/ (accessed 2026-09-03)

[75] Upstarts Media (Alex Konrad), This Startup Wants To Record Your Chores To Train Robots Better, 2025-08-12. https://www.upstartsmedia.com/p/mecka-ai-robotics-data-startup (accessed 2026-09-03)

[76] Intrepid Control Systems (Colt Correa), Time Synchronization in Embedded Systems, overview of gPTP (Precision Time Protocol), webinar video, 42:12, 2021-02-04. https://www.youtube.com/watch?v=ZrujSridNpQ ; slides https://cdn.intrepidcs.net/events/Webinars/Time_Sync_Embedded_Systems_gPTP_20210125.pdf (accessed 2026-09-03)

[77] Intrepid Control Systems, Precision Time Protocol (gPTP) Explained in 5 minutes, video, undated. https://www.youtube.com/watch?v=3vsO4Ndv-9M (accessed 2026-09-03)