Robotics Weekly Analysis — 2026-10-19: Physical Intelligence began selling its models through an API but did not disclose prices

Briefings covered: 2026-10-19 news briefing (period covered 2026-10-13–10-19, with items from 10/7 and 10/12 included retroactively)

Core focus areas: humanoids and Physical AI / drive-train components (actuators, reducers, motors) / industrial robots, cobots and logistics AMRs. Robot brain architectures are divided into three lines — VLA / the world-model approach / WAM (a structure that uses a video world model as the policy backbone) — but we read them not as a contest between camps but as a question of "which layer in a combined stack holds pricing power."

This Week's Key Points

A channel for a model company to collect usage fees directly from customers opened for the first time, but prices have not yet been disclosed. Physical Intelligence (PI), through its "Pi API" beta, lets outside developers fine-tune π0 and π0.5 on their own robot data and run inference in the cloud (10/13). In last edition's Ultra case, the integrator set customer prices; with the API, the model layer sets the unit price directly. Since pricing and GPU allocation have not been disclosed, "which layer holds pricing power" cannot yet be judged. What has become clear instead is what would need to be disclosed to make that judgment.

Industrial safety rules that had blocked Korean humanoids from entering factories were relaxed for the first time through a demonstration exemption. Korea's Occupational Safety and Health Standards assume that industrial robots are fixed and fenced. Under a MOTIE regulatory exemption (10/13), 20 Robros humanoids will go into sites in four industries: semiconductors, automotive, shipbuilding and logistics. In revenue terms it is small, but the institutional precondition for Korean humanoid and drive-train demand has started to move.

Large VC money went not to general-purpose humanoids but to specialized robots with visible revenue. ICON (construction 3D printing, valuation US$1.4 billion, government contracts of US$370 million+), MMI (surgery) and Multiply Labs (biomanufacturing) are examples. In China, there were reports that DeepSeek is leading the seed round of embodied-AI lab Winverse (valuation about US$400 million, not finalized). The picture is one in which in the US, VC money flocks to independent model companies, while in China, LLM labs are expanding vertically into embodied AI.

Whether the video world-model (WAM) approach carries over to real-robot performance again cannot be judged this week. Capital took a step toward video (first-person video data company Mecka AI, US$60 million, led by Sequoia with NVIDIA and Samsung participating), while the evidence is still on the side of field data (this week's only commercial product, Pi API, works only when customers upload their own robot data). Externally reproduced real-robot performance was again zero this week.

Next week brings a cluster of results to check. Tesla's Q3 results (10/21), FANUC's results (10/23) and LG Electronics' Q3 call (late October) are concentrated together.

1. Theme and Value-Chain Implications

(1) Physical Intelligence began selling through an API, and prices are undisclosed

• Confirmed fact (PI official, The AI Insider 10/14): Pi API is a waitlist-based beta. When outside developers upload their own robot data, PI provides hosted fine-tuning and cloud inference. A memory system (MEM) for long-horizon tasks taking more than 10 minutes was also introduced. Initial partners are Actor (autonomous wheel loader) and Almond (dual-arm robot; "learned a deburring policy in four days via the API" is a company announcement). Pricing and GPU allocation are undisclosed.
• Difference from last edition: in the Ultra structure, customers paid the integrator an integration fee and subscription, and PI's share was undisclosed. In the API structure, developers pay usage fees directly to PI. PI now has two revenue paths at once, "via integrators" and "direct API sales," and if unit prices are disclosed, the share the model layer receives in each path can be compared for the first time.
• Reasonable inference: in the API model, customers have to bring their own robot data. This product is not "a general-purpose brain that works without data" but "a service that turns customer data into policies." The bargaining power of hardware and deployment companies that hold site and demonstration data does not disappear immediately. Conversely, PI is positioned to accumulate data feedback through many customers' fine-tuning. Whether its terms allow it to use customer data for retraining is unverified, and this clause is the real variable for pricing power.
• Form factor: neither of the first two customers is a humanoid (wheel loader, dual-arm). It is a sample showing that the first paid demand for robot brains comes from machines whose bodies are already set, rather than from bipedal humanoids. It should not be carried over as a revenue basis for the humanoid theme.
• Classification: PI is in the VLA family using a flow-matching action expert. It is not the world-model approach or WAM.

(2) An exemption was applied to the safety-fence rules that had blocked humanoid demonstrations in Korea

• Confirmed fact (MOTIE press release 10/13): the 6th Industrial Convergence Regulatory Sandbox Committee approved demonstrations of mobile humanoids at manufacturing and logistics sites. 20 Robros humanoids will go into Nepes (semiconductors), Sungwoo Hitech (automotive), HD Hyundai Samho (shipbuilding) and Delivery Lab (logistics). The demonstration period and conditions are unverified.
• Meaning: to put humanoids in a factory, they had to pass the standard for fixed robots surrounded by safety fences. This rule conflicts head-on with the core value of humanoids (deployment directly into human processes). This exemption resolved that conflict at the demonstration level for the first time. Actual process data from four industries will accumulate.
• Reasonable inference: the exemption must lead to a formal revision of the standard for Korean demand to move from demonstration units to mass-production units. Until then, the thesis on Korean revenue for Korean humanoid and component makers is tied to the institutional schedule. Three of the four demonstration sites are listed companies or affiliates of listed groups. In the same direction as last edition's "Hyundai Motor Atlas deployment in 2028," the first source of Korean demand is processes at large automotive, shipbuilding and semiconductor companies.

(3) This week's large rounds concentrated in specialized robots with confirmed revenue

• Confirmed fact (Dealroom weekly tally 10/12): five companies raised US$510 million. ICON Series D US$200 million (co-led by Baillie Gifford and BlackRock, valuation US$1.4 billion, citing CNBC), MMI Series D US$125 million, Multiply Labs Series B US$75 million, Mecka AI, Ultra. In Korea, Hurotics raised a KRW 25 billion Series A (strategic investment by Naver, Daewoong and Lotte).
• Meaning: long-term growth investors have come into vertical robots with a revenue basis such as government contracts and regulatory approvals. This is money on a different axis from the "valuation without revenue" of general-purpose humanoids and the brain layer. All three overseas deals are outside the core focus areas (construction, surgery, biotech), so we do not use them as humanoid valuation references.
• China: DeepSeek leading Winverse's seed round (Bloomberg 10/12, not finalized) is its second robot investment after its existing stake in Unitree (about 2%, per earlier reports). We read it as a picture of LLM labs expanding vertically into embodied AI (reasonable inference, two data points). The model architecture is undisclosed, so we classify it as neither VLA nor world model.

(4) There were no new drive-train component orders, and one case of hydraulic actuators appeared

• There were zero new order disclosures for reducers or actuators in the period. There are also no results yet from the "October big-tech actuator meeting" LG Electronics flagged.
• NeodraDynamics Neo1 (10/14 CIIF): a 4 m, 1.2 t humanoid that drives 39 of its 41 joints with in-house compact hydraulic actuators (company announcement, no independent verification). It has no effect on the logic of reducer demand for human-sized humanoids. We record it only as a note that, in large robots for heavy loads and special environments, hydraulics may be an alternative to "motor + reducer."

(5) No new announcements in the world-model approach, and a related investment in the data layer

• There were no new announcements regarding AMI Labs, General Intuition, 1X World Model or Cosmos Policy. External verification also zero.
• Mecka AI: processes human first-person video and motion into robot training data. It is neither a world model nor a WAM but a data supply layer. However, its investment logic shares roots with the WAM narrative. Sequoia and chip and memory strategic investors (NVIDIA, Qualcomm, Samsung) have bet money on "human video carries over to robot policy." Because the embodiment gap remains, we do not classify it as real-robot action data.
• Counter-evidence: no reports of new adoption of deterministic physics-based approaches (Realtime Robotics, etc.).

WAM approach verdict (this edition): no verdict possible. Investment went into video data (Mecka US$60 million), but this week's only commercial product, Pi API, presupposes customers' site robot data. Externally reproduced real-robot performance: zero.

2. Company Watch — Listed

Tesla (TSLA) — Q3 results on 10/21

• Confirmed fact → on 10/17 Musk said on X that "Optimus and Cybercab are most of Tesla's value." No numerical basis was given. Optimus Gen 3 has no unveiling, specifications or price. The "several hundred units a week production" claim comes from a small blog and is unverified.
• Question to verify → ① Gen 3 unveiling date ② definition of production units (trial production or for external sale) ③ external sales schedule ④ V3 hand design ⑤ mention of Optimus team departures. If the remarks end without numbers, "most of the value" is excluded from the basis for the thesis.

FANUC (6954) — first-half results on 10/23

• Confirmed fact → no order announcements in the period. It lost the No. 1 position in Chinese shipments to local companies last year (reported 9/30). It is building a US$90 million base in Michigan.
• Question to verify → value vs units of robot segment orders in China (whether it is in price competition with local low-cost makers), change in the share of Americas orders. Whether the surge in orders, about which the company itself said it "cannot judge whether it is real or speculative demand," continues.

LG Electronics (066570) — deadline for the outcome of the actuator discussions

• Confirmed fact → it flagged an "October production-readiness meeting on actuators with big tech," but there was no announcement of results as of 10/19.
• Question to verify → whether actuator customers and volumes are mentioned on the Q3 conference call (late October). If there is no mention of a contract or customer by end-October, we treat it as "technical discussion stage."

Regulatory exemption demonstration sites — Nepes, Sungwoo Hitech, HD Hyundai Samho

• Confirmed fact → they became demonstration sites for Robros humanoids. The allocation of units and the period are unverified.
• Question to verify → these companies are buyers, not suppliers, of humanoids. What to check is whether demonstrations convert into purchases (disclosures or reports at the end of the demonstration period). Since an HD Hyundai affiliate is participating, we also look at the relationship with the humanoid internalization plans of HD Hyundai Robotics in the same group (speculation, no supporting evidence).

Angel Robotics (455900) — a private comparable emerges

• Confirmed fact → its private competitor Hurotics raised a KRW 25 billion Series A (cumulative KRW 33.5 billion+). Medical exosuit (MFDS Class 2), US exports in June. Valuation unverified.
• Question to verify → if Hurotics' valuation is disclosed, comparing it with Angel Robotics' market cap would create a Korean wearables pair that allows the gap between listed and private to be measured for the first time.

3. Company Watch — Private and Funding

Winverse raised a seed at a US$400 million valuation without a product. The founder previously led Tencent's vision-language and embodied foundation model team, and the lead investor is China's top LLM lab. With neither revenue nor product, this price is an option price attached to talent and access to a strategic investor (reasonable inference). As it is not finalized, we do not include it in the comparison set.

Mecka presented a revenue-like figure for the first time in the data layer. The company claims a run-rate of US$100 million+. It is annualized, not actual revenue, and there is no external verification. Still, it is a new data-layer sample after XDOF (US$1.2 billion) was dropped from the comparison set for unverified closing. Since the valuation is undisclosed, we do not calculate a multiple.

The gap with listed controls. Unitree has no new price information since -53% from its opening price (9/9). The only private deals with disclosed valuations this week are ICON (based on government contracts) and Winverse (not finalized), and neither is a direct humanoid comparable. Therefore there is no basis for changing last edition's conclusion that "the gap is explained not by revenue but by layer (SW vs HW) and jurisdiction (US private vs China under CSRC review)." We add one point only. Winverse is a candidate sample showing that private prices form for the SW (brain) layer even under Chinese jurisdiction. If finalized, it becomes grounds for lowering the explanatory power of the "jurisdiction" variable and putting more weight on the "layer" variable.

Check date tracking

4. What to Watch

• Disclosure of Pi API unit prices and the clause on retraining with customer data — Why now: a channel for the model layer to collect directly from customers has opened for the first time, but there are no prices. Check: PI's official pricing page and terms of use, the general-availability announcement, follow-up operating metrics from Actor and Almond. If unit prices are disclosed, set them alongside the Ultra path (integrator sets prices) and re-judge "pricing power in the model layer." If the terms allow customer data to be used for retraining, re-record PI's moat not as "the model" but as "the customer-data feedback loop."
• Schedule for a formal revision of the standard following the regulatory exemption — Check: pre-announcement of legislation revising the Occupational Safety and Health Standards by MOTIE and the Ministry of Employment and Labor, notice of the demonstration period, announcements by Robros and the demonstration sites at RoboWorld on 11/4–7. If a legislative pre-announcement appears, fix the start of the thesis on Korean revenue for Korean humanoid and component makers to its effective date.
• Tracking "capital flows (model layer) vs location of the bottleneck (physical layer)" — (a) number of model-layer rounds disclosing revenue or external verification: zero this week (b) overseas mass-production orders for drive-train components: zero. If (a) = 0 for four weeks in a row, record model-layer rounds separately as "capital signals" only.
• Optimus figures on Tesla's Q3 call (10/21), value vs units of FANUC's first-half Chinese robot orders (10/23), actuator mentions on LG Electronics' Q3 call (end-October) — judged under the conditions in Section 2 above.
• Winverse closing and first model release — if confirmed, fix the "LLM lab vertical expansion" sample at two and classify the model architecture according to the company's description.
• The substance of Mecka's run-rate — if two or more customers are confirmed by name, raise it as a candidate reference price for the data layer. If only run-rate claims are repeated, do not use it for multiple calculations.
• FieldAI 11/2 deadline — if neither closing nor a revenue definition is disclosed, delete the multiple (≈74x) and leave it only as a "reported negotiation price."

5. Data Reliability Notes

• Primary sources: MOTIE press release (regulatory exemption), PI official site (Pi API), Dealroom weekly tally (funding).
• Company claims: Almond "learned a policy in four days," Mecka run-rate of US$100 million+, NeodraDynamics dual-arm payload of 500kg.
• Not finalized / single source: Winverse round (one Bloomberg report; other outlets cite it), talk of a Mecka valuation of about US$500 million, Tesla's "several hundred units a week production" claim (small blog), relationship with HD Hyundai Robotics (speculation).

*This article is industry analysis based on public information and is not a recommendation to buy or sell any security.*