Embodied AI Vendor Market Research (Group Overview)

1. Introduction

Embodied intelligence is the main battlefield in the 2026 AI industrialization narrative as it moves from the digital world into the physical world. This group takes "specific vendors' deployment practices" as its unit of study, covering 10 documents and 11 vendors (the 07 combined piece covers three): Unitree, AGIBOT, Figure, Tesla Optimus, Fourier, Galbot, Robotera / Accelerate Evolution / LimX Dynamics (combined piece), Physical Intelligence, UBTech, and the simulation and data platform ecosystem that serves as the infrastructure layer.

Dissecting each vendor as a deployed embodied Harness implementation is what fundamentally distinguishes this group from conventional industry reports: it looks not only at "what the robot can do," but at where in the six-layer model each vendor makes real investment in its model stack, data loop, evaluation system, and governance boundaries.

1.1. Research Scope and Subjects

Selection basis: these 11 vendors cover the main ecological niches of the embodied intelligence industry — embodiment scaling (Unitree), full-stack integration (AGIBOT, UBTech), high-valuation commercialization (Figure), automaker in-house use (Optimus), vertical-scenario differentiation (Fourier, Galbot), new-force route differentiation (piece 07), pure model layer (PI), and infrastructure layer (the simulation piece); each has enough publicly verifiable information to stand on its own.

1.2. Embodied AI Industry Background

According to the Omdia basis cited by Zheshang Securities (2026-03-28), global humanoid robot shipments in 2025 were about 13,000 units, with the leaders being Unitree (over 5,500 units), AGIBOT (5,168 units), and UBTech (~1,000 units); IDC's "Global Humanoid Robot Market Analysis" (2026-01-22) gives an order-of-magnitude figure of ~5,000 units for both AGIBOT and Unitree. The industry broadly agrees that 2026 is the turning year from "demonstration" to "industrialized deployment" (per the WRC 2026 industry observation basis), with three structural forces happening simultaneously:

1. Capital intensification: Figure's Series C valuation of US$39B, Galbot's Series D of RMB 2.5B (the first embodied-AI investment by Big Fund Phase III), LimX Dynamics raising US$400M in six months; Unitree landed on the STAR Market, UBTech is already on the HK exchange, and AGIBOT is advancing on a dual track — 2026 is being called the "IPO year" of embodied AI.
2. Data infrastructuralization: AGIBOT's data collection factory, Beijing's humanoid-robot data super factory (RoboMIND downloads exceeded 20 million), and UBTech's Zigong data collection center have all come online; the industry has about 500,000 hours of high-quality real-robot data vs a demand of ~10 million hours (C-grade basis), making data the #1 bottleneck.
3. Standards and governance getting started: the MIIT's "Humanoid Robot and Embodied Intelligence Standards Framework (2026 Edition)" was released; risk events such as Figure's safety lawsuit and the UBTech U1 controversy turned runtime safety and data governance from a bonus item into an admission requirement.

1.3. Tiered Landscape of Complete Machine / Brain / Embodiment

图 1-1｜整机 / 大脑 / 本体分层格局：厂商生态位全景

数据来源：基于本文分析绘制的示意图。

Unlike the 02-Industry Empowerment/02-Embodied AI group's "general technology stack" perspective (a value-chain decomposition of brain → cerebellum → servo → materials), this group observes the same landscape from the vendor ecological niche:

The key judgment of this landscape is: "the brain" and "the embodiment" are splitting into two distinct businesses. PI's pure-model route, Unitree opening its H2+ reference design to NVIDIA GR00T, and Galbot's heavy bet on algorithms while not self-developing its embodiment (it purchases from Unitree) are all evidence of the same trend. But as of 2026, the most substantive commercialization revenue still sits in the "full-stack integration" and "embodiment scaling" tiers — the pure-brain vendor (PI) has no unit revenue yet, confirming that "before hardware scaling is complete, the model layer's value is hard to monetize independently."

1.4. Anatomy Method

All 10 documents in this group execute the same anatomy framework to ensure horizontal comparability:

2. Glossary

This table is the group-wide baseline for common terminology; each platform document adds vendor-specific concepts on top of it.

3. Market Overview

3.1. Shipping and Commercialization Baseline

Key judgment: the shipping leaders (Unitree / AGIBOT) and the valuation leaders (Figure at US$39B / PI at US$56–110B in talks) are not the same companies — the Chinese market is priced on shipments and scenario validation, while the US market is priced on models and narrative.

3.2. Capital Status Spectrum

3.3. The Data Route Debate

The "route divergence" in embodied AI is essentially a divergence in how data is obtained; in 2026 three routes coexist (see the three-way debate in section 6.3 of 10-simulation-platform.md):

4. Cross-Platform Comparison Matrix

4.1. Basic Attribute Comparison

4.2. Harness Six-Layer Maturity Comparison

Rating notes: ★★★ clear, verifiable mechanisms and evidence; ★★ mechanisms exist but engineering details are not visible or evidence is indirect; ★ public information is missing or relies on a single narrative. Filled in from each volume's "Harness Design" chapter; revisions must update this table in sync.

Horizontal reading: L2 (tool and execution) is the most mature layer across the industry — the engineering dividend of hardware has been fully realized; L5 (evaluation and observation) and L6 (governance and safety) are the common weaknesses — the Harness build-out in embodied AI lags behind the engineering build-out of hardware, mirroring the AI IDE group's judgment that "models are mature but Harness can't keep up."

4.3. Deployment Scenario Coverage Comparison

5. Selection Recommendations

Step one in selection is not comparing specs, but answering three prerequisite questions.

Question 1: Do you want the "embodiment" or the "brain"?

• Need a turnkey complete machine and delivery system → full-stack integrators: for industrial scenarios prioritize UBTech (longest customer list) or AGIBOT (full-stack + open-source ecosystem); for logistics sortation in North America consider Figure (RaaS).
• Already have an embodiment but lack a model → model layer: PI (π series, academic baseline) or NVIDIA GR00T (platform + reference design).
• Need simulation and data infrastructure → 10-simulation-platform.md: choose MuJoCo / Genesis for academia and Isaac for industry; start data public goods with RoboMIND / AgiBot World.

Question 2: What is the constraint function of the scenario?

Question 3: Can you get the data and governance accounts in order?

1. Before any purchase, always ask three things: data route (synthetic / real-robot / feedback), acceptance criteria (require customer-level KPIs written into the contract; refuse acceptance by demo video), safety governance (runtime safety standards are currently industry-wide missing; protect yourself through contract terms).
2. Every shipment and valuation figure must have its basis verified: this group has eight instances of conflicts between two or more bases (see each volume's information-gap declaration); always cite source and reliability when quoting.
3. Multi-robot deployments should agree on operational SLA (following the Galbot × Ning's Home Service 1300+ site model); humanoid robots require significantly higher operational density than traditional automation equipment.

6. Core Judgments of This Group

Judgment 1: The embodied AI of 2026 is an industry where "hardware runs ahead of Harness." The L2 execution layer has ample evidence (31–75 DOF, 3-minute battery swap, 50Hz action chunks), while L5 evaluation (universally missing except for UBTech's audited basis and Figure's operational KPIs) and L6 governance (only three hard pieces of evidence: Fourier's clinical, UBTech's border ports, Galbot's drug license) lag badly. This mirrors the AI IDE group's judgment that "models are mature but Harness can't keep up": every vertical domain is replaying the same scene — capability precedes predictability.

Judgment 2: Shipments and valuation are two separate pricing systems. Unitree / AGIBOT define China's industrial pricing with five-thousand-unit shipments and hundred-million-RMB revenue; Figure / PI define the US market's narrative pricing with ten-billion-dollar valuations. The two systems have not yet converged — PI has no unit revenue, and Figure's valuation corresponds to production that has just crossed one thousand units. The crossing point should appear around 2027: whichever of Chinese vendors' model layer (GO-2 / ERA-42 / COSA) or US vendors' mass-production layer (BotQ / Fremont) closes its gap first.

Judgment 3: Data is currently the #1 bottleneck, and the route has not converged. The gap of 500,000 hours of stock vs 10 million hours of demand (C-grade basis) and the Sim2Real chasm of 89.4% vs 12% determine that the three routes of data collection factory, synthetic data, and deployment feedback will burn money in parallel for the next two years. RoboMIND's 20 million+ downloads show that "data as public goods" is the only validated path to cost reduction.

Judgment 4: Vertical scenario differentiation generates cash flow earlier than general-purpose platforms. The named revenue and compliance access of Fourier (rehabilitation), Galbot (retail / pharmacy), and UBTech (automotive + border ports) all lead the "general-purpose humanoid" narrative vendors. Scenario constraints (human-robot cohabitation, battery-swap continuity, drug regulation) have instead become the moat for product definition.

Judgment 5: L6 is an industry-wide collective gap, and also an opportunity for first movers. Figure's safety lawsuit, UBTech's U1 controversy, and the blank governance disclosures across vendors all mean that standards for runtime safety and data governance have not yet been established. The MIIT standards framework 2026 edition has begun — the vendor that first turns L6 into a product will obtain a positive-sum effect similar to the "sandbox + audit" in the AI IDE market.

7. Division of Labor with the Industry Empowerment Group

The division boundary between this group and the 02-Industry Empowerment/02-Embodied AI group:

Both groups share the Harness six-layer model and the information cutoff basis (2026-09-12); terminology follows Param Card v1.1. When this group's volumes touch technology-stack details (e.g. the whole-body control of WholeBodyVLA, the technical meaning of dexterous-hand DOF), the Industry Empowerment group's corresponding documents are the technical authoritative source; when the Industry Empowerment group touches vendor cases, it cites the fact cards and basis annotations of this group's volumes.

8. Document Navigation

Suggested reading order: first read this README to build a comparison framework, then read the volumes in order 01→10; if you only care about commercialization evidence, prioritize 02, 03, 06, 09; if you only care about model and data methodology, prioritize 07, 08, 10.

9. Summary

The embodied AI market of 2026 presents a landscape of "three layers in parallel and two curves": the embodiment-scaling layer (Unitree / AGIBOT / UBTech) has completed commercialization validation with five-thousand-unit shipments, audited-basis deliveries, and named industrial customers; the model layer (PI, each vendor's VLA / world models) iterates rapidly on methodology but has not yet monetized independently; the infrastructure layer (NVIDIA stack, RoboMIND, data collection factories) is turning data and evaluation into public goods. The two curves are "China's curve, priced on shipments" and "the US curve, priced on narrative," and their intersection is the observation target that all ten documents in this group jointly point to.

The value of dissecting vendors as embodied Harnesses is that it replaces the non-actionable conclusion that "robots aren't good enough" with engineering items that can be addressed one by one: "L5 lacks independent verification," "L6 lacks runtime safety standards," "the L4 data route has not converged." Industry-wide L2 maturity and L5/L6 lag hold simultaneously — the decisive factor of the next competitive cycle (2027) will most likely lie not in degrees of freedom and torque, but in evaluation systems and governance boundaries.

Information Gap Declaration

1. List of conflicts between two or more bases (all presented in parallel in the main text): Unitree 2025 revenue RMB 1.699B / 1.708B; AGIBOT shipments 5,168 units (Omdia) / ~5,000 units (IDC); Figure's return to BMW across three bases (2026-03 / 04 / 06-25 or 07-01); Galbot valuation RMB 20B / 21B / 26B, cargo-pod 170+ / 100+ stores; LimX valuation RMB 15B / 6.8B; UBTech production ramp of 10,000 units / 5,000 units; Optimus's 50 units in Shanghai is only a broker-report basis; PI's Series C US$11B in talks, not closed.
2. Completely missing items: Fourier's shipment volume and financing status; Accelerate Evolution's T1 price and shipments; LimX's named-customer cases; Unitree's H2 pricing; Figure's official RaaS pricing and Catalyst unit count; runtime memory and safety-governance details are broadly missing across vendors.
3. Disallowed materials: the "Anthropic to acquire Physical Intelligence" rumor (D-grade, no first-hand source) was not written into any volume's main text; the Musk remark about Galbot playing tennis (cited without source) was not used.
4. IPO developments (AGIBOT HK, LimX confidential filing, Robotera rumored to sprint toward listing) are all reported / rumored bases and serve only as status notes.
5. This README's six-layer ratings are back-filled from each volume's "Harness Design" chapter; when any volume is revised, the section 4.2 matrix must be updated in sync.

10. References

1. AGIBOT: The full-stack leader in embodied AI (2025 global shipment Omdia baseline / AGIBOT profile) — Zheshang Securities, 2026-03-28. https://www.sgpjbg.com/info/df7be3e7242d311ece3785906662f44c.html
2. WRC 2026 observation: the industry enters the industrialization-deployment phase — Sina Finance / Robot Global News, 2026-09-12. https://www.toutiao.com/article/7684438007058842150/
3. Unitree official website — Unitree, 2026. https://www.unitree.com/?id=1
4. AGIBOT Robotics official website — AGIBOT Innovation, 2026. https://www.agibot.com.cn/
5. Figure official website — Figure AI, 2026. https://www.figure.ai
6. Galbot (Galaxy General Robotics) official website — Galbot, 2026. https://www.galbot.com
7. Robotera official website — Robotera, 2026. https://www.robotera.com/
8. Physical Intelligence official blog — Physical Intelligence, 2024—2026. https://www.pi.website/blog
9. LimX Dynamics official website — LimX Dynamics, 2026. https://www.limxdynamics.com
10. NVIDIA Isaac — AI Robot Development Platform — NVIDIA Developer, 2026. https://developer.nvidia.com/isaac
11. Beijing's "Data Super Factory" helps robots understand the world (RoboMIND 20 million downloads) — People's Daily Online Beijing channel, 2026-09-03. https://bj.people.com.cn/BIG5/n2/2026/0903/c14540-41685467.html
12. UBTech: deep cultivation of core technology, mastering unique skills — Science and Technology Daily, 2026-03. https://www.toutiao.com/article/7613949193794716160/
13. Galbot: founded 34 months, 6 financing rounds totaling RMB 7B — 36Kr, 2026. https://36kr.com/p/3978672823417862
14. Humanoid Robots in 2026 Compared — ValueAdd VC, 2026. https://valueaddvc.com/blog/humanoid-robots-in-2026-figure-apptronik-1x-and-tesla-optimus-compared
15. Embodied AI data surges, but model validation is not yet complete — Sina Finance / Yicai, 2026-09-09. https://k.sina.com.cn/article_1733360754_6750fc7202001in52.html