Embodied AI Daily · September 19, 2026

Today's Highlights

1. Lexiang Technology's "Aether" model livestreams street BBQ in Shanghai — no remote control, no editing, no script: a 4B-parameter model trained on ~200 hours of human video and zero real-robot data. 2. 9th China Robot Summit (Hangzhou) wraps up — with capital hype fading, embodied AI evaluation systems and industry coordination platforms become the concrete deliverables. 3. The data debate at the summit: "Human society never prepared training data for embodied intelligence; there is still no industry consensus on what good data is." 4. Shao Tianlan controversy, act two: his new claim that "intelligent robots are a real trend, humanoids are not" prompts pushback examining his own firm's position as a supplier. 5. Counterpoint: 22,000 humanoid robots shipped globally in H1 2026, +300% YoY — the top five are all Chinese vendors with a combined 86% share.

Details

1. Lexiang's Aether model: an uneditable public verification

• Sources: Xinhua / Sohu Tech / Sina Finance
• Links: Xinhua · Sohu retrospective · Sina Finance
• On September 16, several robots set up grills on a Shanghai street: a BBQ robot tended the grill while multiple service robots took orders and delivered food, with cross-platform relay collaboration — no manual teleoperation, no post-editing, no script. Xinhua says the BBQ work ran for nearly an hour; retrospectives describe the stream as over an hour (one passage says "nearly two hours"; recorded as-is). One tech outlet reported ~5.5 million concurrent viewers (single source). Six registered audience members 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 with no briefing. During the stream, a robot dropped a condiment bottle, picked it back up, and continued.
• Technical claims: Aether has 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 in one energy score — closer to goal and more physically plausible paths mean lower energy; if a glass is out of reach, find a stepping tool), active perception (change viewpoint when information is insufficient), streaming memory (resume task state after interruption), and a Self Model for 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 one-take video circulating in late August, two robots of different brands already shared one model and divided housework autonomously — codenamed MVP ("Make Veritable People"), claimed and renamed Aether by CEO Renjie Guo on September 10.
• Caveat: Lexiang simultaneously published a long article accusing OpenAI's physical-AI piece of plagiarizing its technical roadmap published two months earlier (Evolution Manager vs Recursive Self-Improvement, visual comparisons). OpenAI has not responded. This is a one-sided claim: we report the dispute exists, not that plagiarism is established.

2. 9th China Robot Summit: evaluation systems enter the industry agenda

• Sources: Cailian Press / Xinhua / China Industry News
• Held September 16–18 at Hangzhou Future Sci-Tech City under the theme "Intelligent Connectivity, New Ecosystem — transformation in the Robot+ era." Topics spanned embodied AI / physical AI, humanoid testing and certification, biomimetic and eldercare robots, large models and hardware, across the full chain from core components to capital. Cailian Press's key takeaway: this summit's concrete outcomes were industry coordination platforms and the construction of embodied AI evaluation systems — verification itself became an industry agenda item. Dozens of supply-chain companies exhibited humanoids, quadrupeds, dexterous hands, tactile/force sensors, and joint modules.

3. The data "second half": scarcity, competition, and disputes

• Source: Sina Finance summit coverage, September 18.
• Two industry-level judgments from company representatives: (1) "Human society never prepared training data for embodied intelligence or physical AI; existing stock is small." (2) "The industry has not fully clarified what good data is or what should be collected."
• Context: CAICT's Embodied AI Training Grounds Research Report (2026) estimates only hundreds of thousands to ~1 million hours of high-quality real-world data exist globally — far below what embodied foundation models need for their "ChatGPT moment." Qianche Intelligence's distributed collection network, disclosed in early September, covers 50+ Chinese cities with 100,000+ hours of real-world data in a pretrain–post-train–real-robot validation loop.

4. Shao Tianlan, act two: from attacking others to being scrutinized

• Sources: Sina Finance / Sohu / 36Kr, September 18.
• After his September 9–10 WeChat Moments posts accusing "deal-orchestrating" companies of fabricating unsustainable revenue via related-party data-collection-center transactions (which escalated to a police report by the opposing party, per the Sep 17 digest), Mecamander (Mech-Mind) CEO Shao Tianlan said 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, commentary flipped the lens: Sohu columns titled "Removing the ladder right after IPO" and "Public interest, or protecting the rook / market-cap management" noted that Mecamander itself supplies data-collection equipment, and the attacks objectively serve its own product narrative. The debate is moving from "who is swimming naked" to "where the referee stands."

5. Counterpoint's mass-production stress test: 22,000 units, top five 86% Chinese

• Source: The Paper (citing Counterpoint Research, August 2026 report), September 16: link
• Global humanoid shipments exceeded 22,000 units in H1 2026, up nearly 300% YoY. The top five are all Chinese vendors with 86% combined share:
• Zhiyuan (Agibot): ~9,700 units (43%+ share); revenue from RMB 300k in 2023 to over RMB 1B in a single quarter of 2026; has launched a Hong Kong IPO process targeting a HKD 40–50B valuation.
• Unitree: 7,000+ units (31%); STAR Market debut August 19 with a first-day market cap of RMB 341.8B; H1 revenue RMB 1.152B, net profit RMB 274M (turned profitable); >90% self-developed components; its claim of 5,500+ units shipped in 2025 is self-reported.
• Galaxy General: 1,100+ units; RMB 7B raised, RMB 20B+ valuation; Galbot G1 operating 7×24 on a CATL production line.
• UBTech: thousand-unit scale, full-year guidance raised to 5,000.
• Leju.
• The report projects 50,000+ units for the full year (under 3,000 in 2024). The real exam: mass production with profitability, 2,000 consecutive hours of uptime in factories, and yields that support a ten-thousand-unit BOM.

6. Autos and embodied AI converge: a two-way dependency

• Sources: People's Daily Online / Economic Daily, September 17–18.
• The thesis: "If automakers abandon embodied AI, they will miss a major extension of 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 estimate that ~70–80% of the body-manufacturing supply chain overlaps with NEVs, XPeng Robotics' $900M first round (a record $6.3B post-money), and in-house world models from Geely (WAM) and XPeng (X-World) — automakers are becoming embodied AI's second source of both capital and supply chain.

7. FluxVLA Engine: LimX Dynamics open-sources its VLA engineering stack

• Links: GitHub · Official site
• Open-sourced in April, with dense updates as recently as September 16 and alphaxiv coverage September 15. The humanoid unicorn LimX Dynamics' integrated VLA platform covers the full pipeline: data processing → policy training → simulation evaluation → accelerated deployment → real-robot deployment. 695 GitHub stars and actively maintained; Hugging Face hosts limxdynamics/FluxVLAEngine components (near-zero HF downloads; adoption is mainly code-side). For a Pre-IPO unicorn, exposing the internal stack signals a competitive shift from tool ownership to iteration speed — VLA harness infrastructure is moving from proprietary to shared.

Editorial Observations

Verification formats upgraded this week. Several items point to one word: verification. The Aether livestream effectively outsourced verification bandwidth to spectators — unedited, irreversible, allowed to fail live (the dropped condiment bottle, self-recovered) — a rung on the assertion-strength ladder that demo videos can never reach. Yet it remains a one-sidedly organized event, and "standing up to OpenAI" cannot extend verification credibility to adjudicating a plagiarism dispute. The same week, the Robot Summit wrote "embodied AI evaluation systems" into its outcomes (verification becoming an agenda in itself), and the fuel of the Shao Tianlan controversy is that revenue claims cannot be independently audited. From WRC score cancellations and pre-registered negative results, to street livestreams and IPO-prospectus scrutiny — "proving it runs" is giving way to "proving it's believable," and the acceptance discipline of separating claims from execution is migrating from paper pipelines to product launches and capital markets.

Both ends of the data spectrum appeared the same week. Aether (4B params, 200 hours of video, zero real-robot data) and Zhiyuan's GE-Act 2.0 (30,000 hours of real-robot scaling with no saturation, per the Sep 17 digest) are not a route disagreement but a cost-structure disagreement: one trades architecture for data, the other trades data for architecture. The summit's "no consensus on what good data is" exposes the vacuum in between: with no independent arbiter of data value, there is breeding ground for orchestrated data deals — the economic essence of the Shao Tianlan dispute. Counterpoint's 22,000 units and 86% Chinese share show the "volume" competition has already been settled within China's supply chain; the next axis is intelligence density per unit of data. Watch window: the World Manufacturing Convention in Hefei, September 20–23, as a first test 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 does it, everyone else must explain to the market why they don't.