Embodied AI Daily Digest · 2026-09-19

Today's Highlights

1. LeXiang Technology's Aether large model livestreams street barbecue in Shanghai — no remote control, no editing, no script; 4B parameters, ~200 hours of human video, zero real-robot training data 2. 9th China Robot Summit (Hangzhou) concludes — 'capital noise fades' as embodied AI evaluation system building and industry collaboration platforms become the implementation focus 3. The data debate at the summit — 'Human society never prepared training data for embodied intelligence; there is no industry consensus on what constitutes good data' 4. Shao Tianlan controversy, act two — new claim that 'intelligent robots are a real trend, humanoids are not'; public opinion begins scrutinizing the critics' own supplier positioning 5. Counterpoint: 22,000 humanoid robots shipped globally in H1 2026, +300% YoY — top five are all Chinese vendors with a combined 86% share

Detailed Content

1. LeXiang's Aether Model: An Uneditable Public Verification

• Sources: Xinhua / Sohu Tech / Sina Finance
• Links: Xinhua · Sohu recap · Sina Finance
• Dates: livestream Sept 16; coverage Sept 17–18

On September 16, several robots set up grills on Shanghai streets: a barbecue robot tended the grill while multiple service robots took orders and served food, with cross-embodiment relay collaboration throughout — no human remote control, no post-editing, no preset script. The barbecue work ran for nearly an hour (Xinhua's figure); the full livestream lasted over an hour (recap's figure, though the same article elsewhere says 'nearly two hours' — both recorded as stated). One tech outlet reported ~5.5 million online viewers (single source). Six registered viewers had one job: interfere — change requests on the fly, move objects, interrupt workflows. When one tester was absent, organizers pulled a random spectator from the crowd and had them give commands with no hints. During the livestream, a robot dropped a sauce bottle, picked it back up, and continued.

Technical claims: Aether is only 4B parameters, trained on ~200 hours of human video and 0 hours of real-robot data. The architecture centers on a unified Energy system (task completion and physical constraints folded into one energy evaluation: closer to goal and physically plausible paths mean lower energy — if a glass is out of reach, find a stepping tool; if a stain won't wipe off, infer it's on the back side), active perception (changing viewpoints when information is insufficient), streaming memory (resuming task state after interruption), and a Self Model (internal body representation enabling cross-embodiment adaptation). The design is analogized to a brain–spinal cord hierarchy: high-level understanding and long-horizon planning above, low-level motor control below. In the 10-minute single-take video circulating in late August, two robots of different brands already shared one model and split up room tidying on their own — codenamed MVP (Make Veritable People), claimed and renamed Aether by CEO Renjie Guo on Sept 10.

Caveat: LeXiang simultaneously published a long-form accusation that OpenAI's physical-AI piece 'plagiarized' its technical roadmap from two months earlier (Evolution Manager vs Recursive Self-Improvement, website visual comparisons, etc.). OpenAI has not responded. This is a one-sided claim: this digest accepts that the dispute exists, not that plagiarism is established.

2. 9th China Robot Summit Concludes: Evaluation Systems Enter the Industry Agenda

• Sources: Cailian Press / Xinhua / China Industry News
• Dates: Sept 16–18, Hangzhou Future Sci-Tech City

Under the theme 'Intelligent New Ecosystem, Empowering All Scenarios,' the summit covered embodied AI physical AI, humanoid robot testing and certification, biomimetic robots, elder-care robots, large models and hardware, spanning the full chain from underlying hardware and core models to applications and capital. Cailian Press's on-the-ground report highlighted the key outcome: industry collaboration platform building and embodied AI evaluation system construction — 'verification' itself became an industry agenda item. The expo floor featured dozens of supply-chain companies covering humanoids, quadrupeds, dexterous hands, tactile/force sensors, and joint modules.

3. The Data 'Second Half': Scarcity, Competition, and Controversy

• Source: Sina Finance (summit coverage), Sept 18

Multiple company representatives at the summit offered two industry-level judgments: first, 'human society never prepared training data for embodied intelligence or physical AI; existing data stock is limited'; second, 'the industry has not fully clarified what good data is, or what data should be collected.' Background: a CAICT research report (2026) estimates globally available high-quality real-world data at only hundreds of thousands to a million hours — far below what embodied foundation models need for a 'ChatGPT moment.' Qingshi Intelligence disclosed in early September a distributed collection network covering 50+ Chinese cities with 100,000+ hours of real-world data, now in a pretraining–post-training–real-robot verification loop.

4. Shao Tianlan, Act Two: From Attacking Others to Being Scrutinized

• Sources: Sina Finance / Sohu / 36Kr, Sept 18

After his Sept 9–10 WeChat Moments posts accusing 'deal-making' companies of manufacturing unsustainable revenue via related-party transactions around data-collection centers (escalating to the opposing party filing a police report, per the 9.17 digest), Mech-Mind CEO Shao Tianlan stated at the Tech Innovators Summit: intelligent robots are a real trend, but the humanoid form is not the optimal fit for manufacturing and logistics — solutions must prove themselves within an ROI timeline. The same day, public opinion began reversing scrutiny: Sohu columns titled 'Removing the Ladder Right After IPO' and 'Behind Industry Public Interest: Sacrificing the Rook or Market-Cap Management' noted that Mech-Mind itself is a data-collection equipment supplier, and the attacks objectively favor its own product-line narrative. The debate is shifting from 'who is swimming naked' to 'where the referee stands.'

5. Counterpoint's Mass-Production Test: 22,000 Units in H1, Top Five All Chinese at 86%

• Source: The Paper (citing Counterpoint Research, Aug 2026), Sept 16
• Link: The Paper

Global humanoid shipments exceeded 22,000 units in H1 2026, up nearly 300% YoY, with the top five all Chinese vendors holding a combined 86% share — AgiBot at ~9,700 units (43%+ share; revenue from RMB 300K in 2023 to RMB 1B+ in a single 2026 quarter; HK IPO underway at a target valuation of HKD 40–50B), Unitree at 7,000+ (31%; STAR Market debut Aug 19 with a first-day market cap of RMB 341.8B; H1 revenue RMB 1.152B, net profit RMB 274M turning profitable; >90% parts self-developed; 2025 shipments of 5,500+ are Unitree's own figure), Galbot/Galaxeon at 1,100+ (RMB 7B raised, RMB 20B+ valuation; Galbot G1 running 7×24 on CATL production lines), UBTech at the thousand-unit scale (full-year guidance raised to 5,000), and Leju. The report projects 50,000+ units for the full year (vs. under 3,000 in 2024). The report's closing three questions are the real test: profitability at mass production, 2,000 consecutive hours of uptime on factory floors, and yields that support a ten-thousand-unit BOM.

6. Autos and Embodied AI Converging: A Two-Way Dependency

• Sources: People's Daily Online / Economic Daily, Sept 17–18

The article's two-way judgment: 'If automakers abandon embodied AI, they will miss a major opportunity to extend industrial capability; if the embodied AI industry does not integrate deeply with the automotive technology chain, its industrialization will be slowed.' Combined with the common industry estimate that ~70–80% of the supply chain is shared between robot bodies and NEVs, XPeng Robotics' first-round US$900M raise (US$6.3B post-money, a record single round), and OEM in-house world models (Geely WAM, XPeng X-World) — carmakers are becoming both the second capital source and second supply chain for embodied AI.

7. FluxVLA Engine: LimX Dynamics Open-Sources Its VLA Engineering Stack

• Links: GitHub · Official site
• Dates: open-sourced in April; latest intensive update Sept 16; alphaxiv feature Sept 15

Humanoid unicorn LimX Dynamics open-sourced an integrated VLA engineering platform covering the full pipeline: data processing → policy training → simulation evaluation → accelerated deployment → real-robot deployment. It has 695 GitHub stars and remains actively updated; Hugging Face hosts limxdynamics/FluxVLAEngine components (near-zero HF downloads; adoption is mainly on the code side). For a Pre-IPO unicorn, exposing the internal engineering stack means the competitive axis shifts from 'tool ownership' to 'iteration speed' — VLA harness infrastructure is moving from proprietary to a shared layer.

Editor's Observations

Verification escalates this week. Several events point to the same word: verification. The Aether livestream effectively outsources verification bandwidth to the crowd — uneditable, irreversible, allowing on-the-spot failure (a dropped sauce bottle, picked back up). That is a tier of assertion strength that demo videos can never reach; but it was still a one-side-organized event, and the 'hard challenge to OpenAI' strong-verification cannot spill over into adjudicating the plagiarism dispute. The same week, the Robot Summit wrote 'embodied AI evaluation system building' into its outcomes (verification becoming an industry agenda itself), and fuel for the Shao Tianlan controversy is precisely that revenue claims cannot be independently audited. From WRC result cancellations and ARS pre-registered negative results, to this week's street livestreams and IPO prospectus scrutiny — 'proving it runs' is mass-yielding to 'proving it's trustworthy,' as the assertion–execution separation discipline moves from paper pipelines into launch events and capital markets.

Two poles of the data spectrum appeared in the same week. Aether (4B, 200 hours of video, zero real-robot data) and AgiBot GE-Act 2.0 (30,000 hours of real-robot scaling with no saturation, per the 9.17 digest) — this is not a route disagreement but a cost-structure disagreement: one side trades architecture for data, the other trades data for architecture. And the summit's 'no consensus on what good data is' exposes the vacuum in between: with no independent arbiter of data value, deal-making data transactions have room to grow — that is the economic essence of the Shao Tianlan dispute. Counterpoint's 22,000 units and 86% share show that the competition on 'quantity' has already been fought out within China's supply chain; the next axis is intelligence density per unit of data. Watch window: World Manufacturing Convention in Hefei, Sept 20–23 (first batch of auto × embodied AI landing cases). A falsifiable prediction: within 12 months, 'unedited long-form livestreams' become a standard segment at top vendors' launch events — once the first vendor does it, everyone else must explain to the market why they don't.