Hong Kong Physical AI: Turning Spatial Perception into Commercial Orders

The question this piece answers: for Hesai, Dobot and UBTech, three Hong Kong–listed companies grouped under "physical AI," how far has each one actually gone from product to order, from order to revenue, and from revenue to cash? Data as of the Hong Kong close on October 26, 2026 (captured 8:12 PM ET on October 26; Hong Kong had not yet opened for October 27).

1. A camera, and a rally that came late

On Monday, October 26, Hesai-W (02525) closed at HK$13.57, up 8.56%. Dobot (02432) closed at HK$18.56, up 8.92%. UBTech (09880) rose only 0.44% to HK$68.50. On the Aya heat ranking, Hesai scored 78 and traded 3.3x its normal volume; Dobot's relative volume was higher still, at 4.4x. The ranking used for topic selection recorded Hesai at +9.2%, while the latest Eastmoney/Tencent quote shows +8.56%. The gap is a difference between data sources; the direction of the move is the same.

Many readers will link the rally to Hesai's new product, but the timeline doesn't support that. Hesai launched its "3D world data engine" Kosmo on October 14, together with the 8K AI spatial camera Kosmo One and the Kosmo World cloud platform. Hesai closed at HK$14.29 the day before the launch (10/13). In the three sessions after the launch, the stock fell instead of rising, reaching HK$12.92 on 10/16. On October 22 it dropped another 13.2% in a single day to HK$11.60, and we could not find a verifiable reason for that drop. Even after the 8.56% gain on 10/26, Hesai was still 5.04% below its pre-launch close. A government web page had also described the camera before this, so the 10/26 move can't be read as a reaction to a first-time launch.

Media coverage that day attributed the move to the sector as a whole. Sina Finance, Futu and Tiger all mentioned a cluster of physical-AI research notes over the weekend, J.P. Morgan's sharp upgrade of Unitree, and news of humanoid mass-production milestones, which together lifted humanoid stocks broadly in the afternoon session. Dobot also launched three products at the industry expo, including the VEGA 7 seven-axis force-controlled arm. These explanations come from the media. What we can verify is that many robotics names rose on the same day, so treating the move mainly as sector sentiment is better supported than an "orders expected from Kosmo" story.

Two things to see in this chart: the launch did not lift Hesai, and by 10/26 all three stocks were roughly back to zero. A single big up day does not show any new commercial progress.

The question worth asking, then, is how much of this business has actually been proven, on the path from "we can build it" to "customers keep paying for it."

2. What "spatial perception" actually sells

First, a definition. LiDAR is a laser-ranging sensor. It emits light and times the returns to build a "point cloud" of 3D coordinates. Before a robot can pick up a cup or step around a person, it has to know where those things are in space. That is spatial perception.

Hesai sells three things along this line, and they are at very different stages of commercialization.

The first is the robot's "eyes," the LiDAR itself, and this already sells in volume. According to Hesai's Q2 2026 results (SEC 6-K), robotics LiDAR shipments were 142,371 units, up 193.4% year on year and 22.7% of total shipments of 628,275 units. The company says its JT128 is used by more than 50 embodied-AI companies worldwide, naming Unitree, Galbot and Galaxea among others. A shipment is a completed purchase, so this layer needs no further commercial proof.

The second is Kosmo, the product behind this week's buzz, and it is still at the "launch" stage. Kosmo One combines LiDAR, 8K imaging and a 1-inch CMOS sensor. Per Sina Finance (2026-10-16), it scans an indoor space of about 100 m² in roughly 15 minutes. It uses 3DGS (3D Gaussian Splatting), a technique that rebuilds photos and point clouds into a 3D scene you can view from any angle. Management's example: modeling a single store by hand takes about 300 hours. Note that Kosmo is not mounted on robots. Its job is to "copy" real spaces into digital scenes for robot training, world models, maps and tourism. Its buyers are data and model teams, not robot factories. Partners include Galbot, Dexmal and Amap, which is opening an initial batch of about 1,000 store pilots. None of the public materials we found disclose a price, a ship date or any order value, so on this point Kosmo has not yet made the move from product to order.

The third is robot actuator modules (joint drives). Per 21st Century Business Herald (2026-10-15), Hesai has delivered more than 10,000 sets cumulatively and expects six-figure unit shipments in 2027. That is the company's own forecast, not an achieved number.

Taken together, these three lines give one number we can check: in its Q2 report, Hesai raised its 2026 revenue guidance for "Strategic Growth Initiatives" from RMB100 million to RMB200–300 million. The report does not break out how much of this comes from anything other than robotics LiDAR, so we cannot tell how much Kosmo contributes.

Why would anyone build a separate business around capturing real spaces? Our research corpus offers a mechanism worth borrowing. According to The Robot Report's RoboBusiness agenda guide (2026-10-20), the bottleneck in deploying physical AI is shifting from models to on-site deployment. If a system only works under demo conditions, each new site can take months of engineering rework, turning a single deployment into a "six-figure project." Kosmo's commercial logic is to lower the cost of "learning a new site." The logic holds up, but the only support so far is the company's own example (for instance, Sharpa robots reaching a 99% task success rate in an ice-cream store), with no third-party quantitative data yet.

3. Key question one: have customers moved from trial to purchase?

The most direct evidence that a robot customer is really buying comes down to three things: they paid and the seller recognized revenue, their volume is above prototype level, and they bought again. Applying those three tests gives a different answer for each company.

Dobot has many customers, but embodied-AI revenue is still small in absolute terms. According to China Fortune's report on Dobot's interim results (2026-08-26), Dobot has 231 embodied-AI customers in total, nearly 100 of them in industrial manufacturing. H1 embodied-AI revenue exceeded RMB45 million, up more than 20x year on year and about 14.2% of total revenue of RMB316 million. The company says its customer pool "covers every commercial conversion stage, from order signing and sample validation to small-batch delivery and large-scale use." However, we could not get a verifiable breakdown of how many customers sit at each stage. A widely circulated split of "172 in sampling, 30 in batch purchase" could not be confirmed in the filings, so we do not use it. Dobot's real revenue base is still collaborative arms: industrial-manufacturing revenue was about RMB200 million, up 142.4%.

UBTech is already recognizing humanoid revenue, and the amounts are meaningful. Per Eastmoney (2026-08-29), H1 revenue from full-size humanoids was RMB590 million, up 1,445%, on 921 units; about 600 Walker S2 units were delivered in H1. Dividing RMB590 million by 921 units gives roughly RMB640,000 per unit. That figure blends different models, so treat it only as an order of magnitude. Humanoids made up 46.5% of total revenue, which makes them UBTech's largest revenue line.

Hesai sits further upstream. Its purchase evidence is the strongest, but each sale is the smallest. Every robotics LiDAR shipped is a transaction. But many of its customers are robot makers still in R&D or small-batch production, so selling a lot of LiDAR does not mean those customers' robots are running at scale in factories.

The counter-evidence belongs here too. The Robot Report (2026-10-11) argues that real commercialization of physical AI should be judged by mean time between interventions (MTBI), cost per task, time to redeploy to a new task, and safety incidents per operating hour. None of the three companies discloses these field metrics. Revenue recognition only proves that robots were sold, not that they keep working at the customer's site. The second point is what determines whether customers buy again.

4. Key question two: why don't orders, deliveries and revenue match?

An order is not revenue, and this is most visible at the robot makers.

UBTech booked more than RMB1 billion of new orders in H1, including three worth over RMB100 million each. According to late-2025 reports from Jiemian, IThome and others, cumulative orders since the start of 2025 were close to RMB1.4 billion. Full-size humanoid revenue recognized over the same period was RMB590 million. Several explanations for the gap can coexist. First, ordinary timing: revenue is recognized only after robots are commissioned and accepted at the customer's site. Second, some orders are framework agreements or staged deliveries. Third, the nature of the orders. The Fangchenggang (Guangxi) contract won in November 2025, for example, was worth RMB264 million and covered a "humanoid data-collection and testing center" plus an AI education demonstration project. That is government-led data and training infrastructure, which is not the same thing as a factory buying robots to replace labor. The available material cannot tell us how much each explanation accounts for.

Two balance-sheet signals stand out. According to interim-report summaries from Ruicaijing and other outlets, UBTech's H1 inventory was RMB985 million, up 70.93% from end-2025. Credit impairment was RMB91.07 million, of which receivables bad-debt losses were RMB76.78 million, against only RMB1.30 million of credit impairment a year earlier. Rising inventory can be read as building stock for H2 deliveries. The sharp rise in bad-debt provisions suggests some customers are paying more slowly than expected. Both sit at the step where orders turn into revenue, not at the product-development step.

For Dobot, we could not verify a backlog figure. CITIC Securities forecasts 2026 revenue of RMB762 million. Working backward, that requires about RMB446 million in H2, 41.1% above H1. This is a broker forecast, not company guidance.

Hesai has the shortest path from order to revenue. Its Q3 revenue guidance is RMB1.10–1.15 billion; the midpoint is 30.7% above Q2's RMB860.8 million. That growth still comes mainly from automotive ADAS (485,904 ADAS units shipped in Q2), with robotics acting as an amplifier.

The table below puts the three companies on the same basis (H1 2026, RMB 100 million).

H1 2026	Hesai	Dobot	UBTech
Revenue (YoY)	15.41 (+25.1%)	3.16 (+106.6%)	12.69 (+104.2%)
Gross margin	39.7%	47.4%	44.7%
Net income attributable	+0.89	-1.06	-3.11
Verifiable robot revenue/volume	Q2 robotics LiDAR 142k units	Embodied AI >0.45	Full-size humanoid 5.9 (921 units)
Operating cash flow	Not obtained	Not obtained	-7.49

5. Key question three: two business models, two cash curves

A component supplier selling "eyes" and a company building whole robots burn cash in very different ways.

Hesai sells shovels. It sells LiDAR to more than 50 embodied-AI companies and does not need to bet on which robot maker wins. As of June 30 it held RMB7.05 billion in cash, equivalents, restricted cash and short-term investments. It has posted five straight quarters of GAAP profit (net income under US accounting rules), with Q2 net income of RMB70.6 million. The trade-off is low unit prices and price-sensitive robot customers. Most of its revenue also still comes from cars, so robotics alone can hardly drive the stock.

UBTech builds whole robots, and its cash burn is much heavier. H1 loss for the period was RMB339 million, but operating cash outflow was RMB749 million, 2.21x the loss. The difference was mostly tied up in inventory and receivables, consistent with the balance-sheet signals in the previous section. Over a longer period, operating cash flow was -RMB1.0 billion, -RMB884 million and -RMB784 million in 2023–2025. Since listing, UBTech has done six share placements raising more than HK$7.4 billion in total. A robot maker has to carry its own inventory, on-site commissioning and customer payment cycles, and each of these ties up cash. A corpus piece (Asianometry, 2020-12-08) made a similar observation about AI companies: AI capability is the way in with customers, but large revenue comes from delivering hardware along with it, which pulls the business model away from replicable software toward shipment-bound, customized system integration.

Dobot sits in between. Its collaborative arms are a mature product with a 47.4% gross margin. H1 net loss attributable was RMB106 million, and R&D spending was RMB102 million, or 32.1% of revenue, so the loss mainly reflects investment in embodied AI. We could not obtain Dobot's period-end cash or operating cash flow from a verifiable source this time, so this part is kept brief.

The chart shows that UBTech's revenue is now close to Hesai's, but its profit runs the other way. Even with revenue growing, the robot-maker model is still consuming profit.

6. What the market is betting on

The price-to-sales ratio (PS: market cap divided by trailing 12-month revenue) gives a rough read on how much the market pays for each unit of revenue. Using 10/26 closing market caps and assuming HK$1 ≈ RMB0.87 (derived from the USD conversion rate in Hesai's filing and the HKD peg), Hesai trades at about 4.5x, Dobot at about 10.8x and UBTech at about 11.3x. Tencent's quote shows Hesai's trailing P/E at 35.5x.

What this chart shows is that the valuations of the two robot makers rest on a bet that humanoids become a very large market and that most of the value stays with the companies building them.

That implied expectation splits into two halves.

For Hesai, the market is essentially pricing a profitable automotive LiDAR company with a robotics option attached. Its cash is about 47.4% of its market cap in RMB terms. The midpoint of the RMB200–300 million Strategic Growth Initiatives guidance is only about 8.3% of 2025 full-year revenue of RMB3.03 billion. Even if Kosmo succeeds, it is unlikely to change the revenue mix in the near term. In other words, Hesai's stock is not very sensitive to Kosmo, which fits the flat-to-down move after the launch in Chart 1.

For UBTech and Dobot, the market is paying in advance for pilots turning into repeat purchases at scale. The evidence for this pricing is real revenue growth (both doubled revenue in H1) and orders above RMB100 million. The evidence against is negative operating cash flow, rising inventory and bad debts, and, as noted above, the absence of field-operation metrics. Our view: for the robot makers' high valuations to hold, H2 needs to show two things. Orders have to convert to revenue on schedule, and receivables have to be collected. There is not yet enough data on either.

7. What to watch next

Hesai reports Q3 results on November 10, 2026 (7:00 AM ET call). This is the nearest and firmest test. Watch four things: whether revenue lands within the RMB1.10–1.15 billion guidance; whether robotics LiDAR shipments keep triple-digit year-on-year growth; whether the RMB200–300 million Strategic Growth Initiatives guidance is raised again; and whether management gives Kosmo's price, ship date or first orders for the first time. If the fourth item is still missing, Kosmo stays a "product" and not yet an "order."

UBTech's H2 deliveries and collections. Most Hong Kong main-board companies report only interim and annual results, so UBTech's next full report is expected around the end of Q1 2027 (check the company's announcement for the date). Until then, watch whether newly announced large orders are factory purchases or data-center and demonstration projects, and track Walker S2 delivery milestones. At the annual results, the key items are whether inventory comes down and whether bad-debt provisions keep growing.

Whether Dobot reaches the roughly RMB446 million of H2 revenue implied by broker forecasts. The key points are whether embodied-AI revenue can rise from 14.2% of the total, and whether the company discloses verifiable funnel data such as the number of customers buying in batches.

8. Conclusion

Within Hesai's spatial-perception business, robotics LiDAR already sells in volume, with 142,371 units shipped in Q2. The Kosmo spatial camera at the center of this week's buzz, however, has no disclosed price, ship date or orders and is still at the launch stage; the 10/26 rally looks more like sector sentiment than a valuation of Kosmo. On the robot-maker side, UBTech recognized RMB590 million of humanoid revenue in H1, so orders have started turning into revenue, at the cost of RMB749 million of operating cash outflow plus rising inventory and bad debts. The market pays about 2.5x more per unit of revenue for the robot makers than for the supplier (about 11x vs. 4.5x). That is a bet on repeat purchases at scale, which Hesai's November 10 results and the two robot makers' annual reports will help test.

Appendix: Data and Sources

A. Key financials and valuation (RMB million unless noted)

Metric	Hesai	Dobot	UBTech	Formula / inputs
FY2025 revenue	3,027.6	492.2	2,001.0	eastmoney.financials
H1 2025 revenue	1,231.7	153.1	621.5	same
H1 2026 revenue	1,541.4	316.3	1,269.1	same
TTM revenue	3,337.3	655.4	2,648.7	FY2025 + H1'26 − H1'25
HK market cap 10/26 (HK$100m)	171.14	81.66	344.83	Tencent quote
Market cap in RMB (100m, est.)	148.8	71.0	299.9	×0.8696 (= 6.7833/7.8; 6.7833 = Hesai 6-K rate 860.8/126.9; 7.8 = HKD peg midpoint)
PS (TTM, est.)	4.5x	10.8x	11.3x	RMB market cap / TTM revenue
Hesai cash / market cap	47.4%	—	—	70.506 / 148.82
UBTech OCF / period loss	—	—	2.21x	7.49 / 3.39
UBTech avg. full-size humanoid price	—	—	~RMB640k/unit	590m / 921
Robotics share of Q2 shipments	22.7%	—	—	142,371 / 628,275
Q3 guidance midpoint QoQ	+30.7%	—	—	1,125 / 860.8 − 1
SGI midpoint / FY2025 revenue	8.3%	—	—	250 / 3,027.6
Dobot implied H2'26 revenue	—	RMB446m (+41.1% vs H1)	—	CITIC 762 − 316

Note: eastmoney.financials labels the currency as HKD, but all three companies report in RMB, and the figures match the RMB numbers in their filings (e.g., Hesai Q2 = 1,541.4 − 680.6 = 860.8, consistent with the 6-K). This report treats them as RMB.

B. Timeline

• Nov 2025: UBTech wins RMB264m Fangchenggang data-collection center and related projects (media).
• 2026-08-24 / 08-28: Dobot / UBTech interim results (HKEXnews).
• Aug 2026: Hesai Q2 results: 142,371 robotics LiDAR units; SGI guidance raised to RMB200–300m (SEC 6-K).
• 2026-10-14: Hesai launches Kosmo / Kosmo One / Kosmo World.
• 2026-10-22: Hesai HK shares fall 13.2% (cause not verified).
• 2026-10-26: Robotics stocks rally; Hesai +8.56%, Dobot +8.92%.
• 2026-11-10: Hesai Q3 results and call.

C. Aya heat data (HK list, 2026-10-26 close)

• 02525 Hesai-W: #6, heat 78, +9.2%, rel. volume 3.3x
• 02432 Dobot: #12, heat 69, +8.9%, rel. volume 4.4x
• 09880 UBTech: #14, heat 68, +0.4%, rel. volume 2.7x

D. Market views (authors' own)

• X @Deelu179242 (2026-10-22): UBTech "first to deliver full-size humanoids at thousand-unit scale"; Dobot's industrial ATOM "has begun its third batch of mass-production deliveries."
• X @SYU77996 (2026-10-24): humanoids + physical AI are moving "from proof of concept to proof of mass production," a 0-to-1 stage.
• CITIC Securities (via Gelonghui): Dobot 2026–28E revenue RMB762m / 1,131m / 1,607m; net income −140m / −62m / +84m; Buy.
• Daiwa (10jqka headline, 2026-09-10): UBTech commercialization momentum strengthening; reiterates Buy.

E. Sources

Tencent Securities quotes and daily bars (2026-10-26); Eastmoney financials; Hesai Form 6-K Q2 2026 (sec.gov); Hesai Q3 date 6-K (sec.gov); Sina Finance 2026-10-16; 21st Century Business Herald 2026-10-15; China Fortune 2026-08-26; Gelonghui; Eastmoney 2026-08-29; media summaries of UBTech interim (Aug–Sep 2026); Sina/Futu/Tiger 2026-10-26; The Robot Report 2026-10-20 and 2026-10-11; Asianometry 2020-12-08; X via grok-x search.

For information and research discussion only. Not investment advice.