Mecka AI Alternatives: A Map of the Physical AI Data Engine Category

Why this comparison exists

The "data engine for physical AI" category is roughly eighteen months old. In that window, more than $570M of disclosed venture and growth capital has been deployed into companies whose job is to produce, label, or broker the demonstration data that robotics foundation models need. The category did not meaningfully exist when Mecka AI was founded in 2025. Today there are at least sixteen credible vendors, three distinct collection models, and zero neutral comparisons published publicly.

We are Kindly Robotics. We are pre-seed, pre-revenue, pre-contract, and we are building a clinical-vertical data engine — which means we have an obvious bias in writing this post and you should read it accordingly. Every funding number, customer name, and date below comes from the company's own announcements or named press coverage logged in our internal database as of June 8, 2026. Where claims are unverifiable, we say so. The Kindly pitch lives at the bottom, separated by a horizontal rule.

The three approaches

The category sorts cleanly into three collection models. They are not mutually exclusive — Lightwheel and Sensei both straddle two — but a vendor's primary motion almost always lands in one bucket.

Egocentric human

A camera (head-, chest-, or wrist-worn) plus, in the richer rigs, body pose sensors, tactile gloves, and force-torque sensing. A human performs an everyday task; the output is first-person video plus aligned kinematics that a VLA model can ingest as demonstration. The pitch: humans scale linearly with population, robots scale linearly with hardware capex. The eight egocentric players we track are Mecka AI, Build AI, Cortex AI, Lightwheel EgoSuite, Awign, Human Archive, Neocambrian AI, and Luel. Mecka's EgoVerse paper (arXiv 2604.07607, April 2026) is the closest thing the category has to a foundational reference text — 1,362 hours, 80,000 episodes, 1,965 tasks, co-authored with Georgia Tech, Stanford, UCSD, ETH Zurich, MIT CSAIL, Meta Reality Labs and Scale AI.

Teleop-as-a-service

A trained operator drives a real robot remotely; every session is, by construction, a demonstration. Differentiators are usually latency (Adamo claims sub-40ms), hardware cost (Sensei's sub-$300 exoskeleton), or vertical depth (Cogito's leased-line infrastructure for surgical and EOD applications). The four teleop-infra plays are Sensei Robotics, Adamo, Cogito Tech, and Extend Robotics. The honest version: this layer is commoditizing — latency budgets compress yearly and the moat is increasingly enterprise distribution.

Platform and annotation

Bring your own data, pay for the lifecycle infrastructure underneath: ingestion, versioning, multi-modal annotation, eval, governance. The three serious entrants are Encord, Scale AI, and Surge AI. None collect data first; all can if a customer pays for it. Scale's UR AI Trainer launch with Universal Robots at GTC March 2026 is the first public signal that a platform incumbent will operate physical collection cells directly. Encord's customer data under management grew from 1PB to 5PB in a year, physical-AI revenue up 10x.

The funded players, by funding stage

This section is short profiles only. The numbers are pulled from each company's own announcements or the named press cited in our database.

Tier 1 — $100M+ disclosed

Lightwheel AI ($140M, multi-round PE/VC, Shanghai). Founded 2023 by a former NVIDIA simulation head. Investors include Amber Capital, Jiupai Capital, and 37 Interactive Entertainment. EgoSuite collects 20,000 hours per week per company blog. Customer list is the broadest in the category: AgiBot, BYD, ByteDance, Figure, Fourier, Galbot, Geely, Google DeepMind, Zordi, plus NVIDIA as co-development partner. $100M in Q1 2026 orders disclosed via PRNewswire — the only nine-figure order book confirmed in the space.

Encord ($110M total, $60M Series C February 2026 led by Wellington Management, $550M valuation, London). Founded 2020. Named physical-AI customers include Woven by Toyota, Zipline, Skydio, and AXA Financial. Platform serves 300+ AI teams.

Cogito Tech (~$159M cumulative pre-acquisition, acquired by Verint October 2024, East Brunswick NJ). RoboStream platform for high-stakes teleop with ISO 9001/27001 and SOC 2. Customers undisclosed — a recurring pattern at the enterprise tier.

Tier 2 — $20M to $100M

Mecka AI ($68M total: $8M seed led by Neo in 2025; $25M Series A part one led by Framework Ventures Nov 2025; $35M Series A part two led by Framework Ventures Q1 2026, NYC). Forty employees. Only confirmed customer is 1X Technologies. CEO has publicly claimed $100M in signed contracts as annualized run rate, with additional Fortune 100 customers unnamed. EgoVerse academic dataset is 1,362 hours; commercial Egoverse is claimed at 30,000+ hours.

Luel ($31.7M seed, 2026, co-led by General Catalyst and Lightspeed, YC W26). Rights-cleared multimodal marketplace — robotics is one of several verticals alongside computer-use, multilingual speech, and imagery. Active projects include gemstone carving, cooking, household chores, warehouse ops. No named customers public.

Sensei Robotics ($23.1M aggregate, BlueCrow + Kamay + Iberis + Seaya, YC S24). Sub-$300 sensorized exoskeleton arm plus operator marketplace. Targets warehousing, logistics, last-mile, and hospitality. Revenue estimate $3.9M per Extruct AI. No named customers public.

Tier 3 — under $20M

Human Archive ($8.2M seed May 2026, Wing VC lead, with NVP Capital, YC, and angels from OpenAI / NVIDIA / Meta). 1,000+ active headset units in India. Hardware stack is among the richest in the category — camera caps, tactile gloves, full-body mocap, wrist cameras, synced with RGB-D. No named customers public.

Cortex AI (~$6M seed 2025, YC F25, San Francisco). Three employees. Founded by Carousell co-founder Lucas Ngoo. Two-sided marketplace where workplaces earn revenue hosting collection sessions. Customers described as "frontier labs" but unnamed.

Build AI ($5M seed 2025, Abstract Ventures, Pear VC, HF0, ZFellows). Solo founder Eddy Xu (Columbia dropout). Open-source: Egocentric-10K, Egocentric-100K, and Egocentric-1M on Hugging Face, with the 1M-hour release in April 2026. No named customers, no disclosed revenue — distribution-first, monetization-later.

Extend Robotics ($0.41M seed, Henley Business Angels, UK). AMAS XR teleop certified through the Universal Robots UR+ ecosystem, distributed in the UK by RARUK. arXiv 2506.01135 is their published work on network latency in XR teleop.

Undisclosed funding

Adamo (pre-seed, F4 Fund). Claru (Moonvalley spinout, no public number; Moonvalley itself raised a $70M seed from General Catalyst, Khosla, and Bessemer). Neocambrian AI (pre-seed completed, undisclosed, India). Awign acquired by Mynavi for an undisclosed amount; pre-acquisition raised roughly $15M per Crunchbase and now operates 1.5M+ Indian gig workers.

Bootstrapped or strategic

Surge AI is bootstrapped, $1B+ revenue in 2024, with a reported ~$25B private valuation per the West Operators case study. Customers are OpenAI, Google, Microsoft, Meta, and Anthropic — all LLM workloads today, latent physical-AI threat tomorrow. Scale AI is a different beast: Meta took a major stake in 2024, valuation widely reported at $14B+, physical AI is a vertical inside the larger company. Named physical-AI customers include Physical Intelligence, Generalist AI, Cobot, and Universal Robots.

When customers are unnamed across nearly every Tier-2 and Tier-3 vendor, the appropriate read is skepticism. "Signed Fortune 100 contracts, names confidential" is the standard PR posture; it is also unfalsifiable. The only company in this category with publicly verifiable blue-chip customer concentration today is Lightwheel.

What's missing from the category

The egocentric data being collected today clusters into a narrow set of environments: factory floors, warehouses, kitchens, baristas, sneaker resale, household chores. That maps to early customer demand from humanoid OEMs targeting general-purpose home and industrial work. Fine.

The verticals conspicuously absent:

• Clinical workflows. No vendor in our database is collecting consented, IRB-cleared data from operating rooms, sterile processing, infusion suites, or bedside care. The data is enormously valuable to the medical-robotics and surgical-assistant model space and is gated behind regulatory infrastructure that consumer-grade gig collection cannot touch. Cogito Tech's surgical-teleop pitch is the closest existing posture — and that is teleop, not egocentric.
• Defense and dual-use. EOD, ISR, and contested-logistics demonstration data is collected today inside DoD primes and cleared contractors. None of it flows into the open egocentric corpus. Cogito's EOD vertical and the DARPA pedigree of Sensei's founders are adjacent signals; neither is a defense-cleared collection program.
• Pharma manufacturing. Aseptic fill-finish, isolator work, cleanroom material handling — high-value, highly regulated, and not legally collectable by handing a head-mounted camera to a contract gig worker.

The pattern is the same in each case: the data is gated by a legal, regulatory, or cleared-environment posture that the dominant egocentric model cannot acquire by scaling consumer gig labor. This is not a claim that Kindly has solved it — only that the gap exists and that no funded competitor we can identify is targeting it as their primary motion.

How to evaluate a vendor in this category

If you are an OEM or foundation-model lab buying from this category in the next twelve months, these are the questions we would want answered before signing.

1. What is the actual recipe of one hour of your data? How many minutes of demonstration, captured with what sensor stack, at what resolution and frame rate, with what sync tolerance between video and pose, with what metadata schema. Anyone who answers in hours instead of in spec is selling you a number, not a dataset.
2. What is your per-hour effective dollar rate, all-in? Not the operator wage. Include rig amortization, QA, annotation, legal review, and vendor margin. The spread across this category is at least 20x today.
3. What is your IRB, consent, and rights-clearance posture? Specifically: written contributor consent, jurisdiction of capture, image-use license terms, and downstream-training rights. Several vendors are operating on standard ToS that may not survive enterprise legal review and certainly will not survive clinical or defense buyers.
4. What is your moat in 24 months when OEMs internalize collection? 1X already runs NEO Expert Mode; Figure has Project Go-Big. Every OEM-tier customer is one engineering hire away from doing it themselves.
5. Who owns the data after I license it? Exclusive, semi-exclusive, or pooled? Can the vendor resell the embodiment-matched portion to your competitor? If the vendor folds in 18 months, where does the dataset go?

A serious vendor will have a one-paragraph answer ready for each.

---

Disclosure

Kindly Robotics is a pre-seed, pre-revenue, pre-contract company. We are building a data engine focused on clinical workflows — a vertical we believe is structurally underserved by the egocentric players profiled above. We have no customer logos to disclose, because we have no customers yet. We have no funding round to announce, because we have not raised one. We wrote this post because we spent the last several months building an internal map of this category for our own use, and a neutral version did not exist publicly. The full pitch — including how we think a clinical wedge fits next to the eight egocentric incumbents — lives at /pitch. If you are a buyer evaluating any of the vendors above and think we have gotten something wrong, we want to hear about it.