Two Hong Kong-listed robotics companies submit their annual reports: Embodied Intelligence is still burning cash, and industrial scenarios are key to breaking the deadlock

Embodied AI is still in the investment phase, but the growth in revenue also reflects that the deployment of humanoid robots in real-world applications is accelerating.

In a results briefing this afternoon, Unitree (09880.HK) management disclosed that the company expects to deliver 5,000 full-size embodied AI humanoid robots in its Walker S series in 2026. The company’s overall capacity planning for full-size embodied AI humanoid robots is expected to exceed 10k units.

The industry is looking forward to an application boom for the embodied AI industry, represented by humanoid robots, and has also shared related outlooks. In the past two days, two Hong Kong-listed embodied AI companies, Abyss (02432.HK) and Unitree, have consecutively released their financial reports, delivering their latest answers on commercialization. The financial reports show that both companies’ revenue increased last year, but their loss situation still needs to be turned around. The revenue share contributed by full-size embodied humanoid robots from both companies has not yet exceeded half. At the same time, R&D investment in embodied AI is still increasing.

From the financial report data, embodied AI is still in the investment phase. However, the growth in revenue also reflects that the deployment of humanoid robots is accelerating.

Embodied AI revenue share is still not high

In 2025, Abyss’s revenue was 492 million yuan, up 31.7% year over year; it recorded a loss of 84.05M yuan, with losses narrowing by 11.32M yuan year over year. Unitree’s revenue last year was 2B yuan, up 53.3% year over year; it recorded a loss of 790 million yuan, with losses narrowing by 370 million yuan year over year.

Abyss is the “first collaboration robot stock” in Hong Kong. Last year, it released products including the humanoid robot Dobot Atom and the wheeled humanoid robot Dobot ATOM-W, entering the embodied AI field from the collaborative robotics segment. Unitree, meanwhile, is China’s “first humanoid robot stock.” The performance of the two companies, to a certain extent, reflects the progress of commercialization for embodied AI.

In terms of product structure, full-size embodied AI humanoid robot-related products are the fastest-growing segment in Unitree’s revenue, but the revenue share has not yet exceeded half.

Last year, revenue from Unitree’s full-size embodied AI humanoid robot products and solutions was 820 million yuan, up 22 times year over year, accounting for 41.1% of total revenue. Revenue from non-embodied AI humanoid robot products and solutions, other intelligent robot products and solutions, and other intelligent hardware equipment together accounted for 58.7%.

Abyss’s entry into the embodied AI field came relatively later. Last year, revenue from six-axis and four-axis collaborative robots and composite robots accounted for a total of 94%. Revenue from embodied AI robots was 20.04M yuan, accounting for 4.1%. However, embodied AI robot revenue is growing rapidly; it increased by more than 4 times last year.

Both companies have投入 considerable R&D resources into embodied AI. Abyss’s R&D investment is growing quickly. Last year it投入 115 million yuan, up about 60% year over year; the increase mainly went to embodied AI. Unitree’s R&D expenditure last year was 508 million yuan, up 6% year over year, of which 270 million yuan was used for the R&D of full-size embodied AI humanoid robots. Unitree’s R&D expenditure as a percentage of revenue decreased from 36.6% in 2024 to 25.4% in 2025. Abyss’s R&D expenditure as a percentage of revenue was 23.3% last year, up 4 percentage points year over year.

The investment in embodied AI is the reason Abyss is still making losses. If you exclude R&D investment of more than 1.84B yuan into embodied AI, the company basically achieved breakeven last year. Abyss management said that, at a normal growth pace, its Collaborative Robotics business unit can maintain profitability this year, giving the company more room to invest in robot intelligence R&D.

As for Unitree, which entered embodied AI earlier, while the increase in humanoid robot sales brings revenue growth, its accounts receivable balance remains high. Abyss’s trade receivables last year were 112 million yuan, while Unitree’s accounts receivable reached 10k yuan, up 40% year over year. Unitree’s accounts receivable last year exceeded nine-tenths of revenue, and its allowance for bad debts reached 539 million yuan, up 140 million yuan from the previous year. Of that, accounts receivable with an aging of more than three years were 342 million yuan, up 260 million yuan from the previous year.

Only when it takes root in industry can the trillion-level market be truly unlocked

Both of last year’s publicly listed embodied AI companies were pushing humanoid robot deployments in real applications.

Last year, Abyss rolled out multiple humanoid robots as well as quadruped robots and six-legged bionic robots. The company said humanoid robots have been deployed in scenarios including the automotive industry chain, commercial services, scientific research, and education. Unitree’s sales of full-size embodied AI humanoid robots rose from single-digit numbers in the prior year to 1,079 units last year, with more than 80% of sales coming from industrial scenarios.

What new developments are there in the embodied AI market represented by humanoid robots? At the performance briefing held after the financial report releases, both Abyss and Unitree management offered their judgments.

“A decade ago, robots grew eyes; ten years later, robots grew brains.” Abyss founder and CEO Liu Peichao said that the embodied AI market began gradually rolling out in the past year, and this year will be the first year for deploying in vertical scenarios.

Similar to how Unitree was the first to push deployment of humanoid robots in industrial scenarios, Liu Peichao also believes that, in sequence, embodied AI will first take root in industrial and commercial sectors, and then move into the final battlefield market—home services.

“To G (for government) and research are last year’s main markets. The second step will enter To B commercial and industrial markets.” Liu Peichao said that last year, the motion performance of robots such as running and jumping basically reached the boundary for practical deployment. Further development will enter interaction scenarios in production as well as in-home or service settings. In industrial scenarios, embodied AI robots can already complete component搬运 (part handling). The company has received some orders and expects that some simpler operational steps will gradually roll out starting in the second quarter. Only after embodied robots achieve breakthroughs in operational capabilities in industrial scenarios will a trillion-level market be truly opened.

In terms of form factor, Liu Peichao believes that judging by the operational capabilities of full-size humanoid robots, the future share of bipedal and wheeled humanoid robots may be 1:3. For quadruped robots, demand in overseas security markets is already clear; in addition, there are also markets for consumer companionship and sports/fitness interaction. He believes that robots that enter homes first in the future will also be quadruped robots. He expects the company’s quadruped robot shipments to be at the ten-thousand-unit level within 1 to 2 years, and the promotion target for this year is around 10,000 units.

After breakthroughs are achieved in robots’ running and jumping capabilities, how a robot’s “brain” works and how to break through in that area has become a focus in the industry.

Liu Peichao mentioned issues related to models and data for embodied AI robots. He said that in factory scenarios, models and data with good consistency can be deployed onto robots within a day, while those with poor consistency are essentially equivalent to requiring multiple rounds of R&D.

“Valuable data accumulated within the industry last year—(duration) is no more than 30k hours. There may be data amounting to tens of thousands of hours that is not very valuable; it can only be used for pre-training, and it is difficult to achieve generalization and improve (robot operation) accuracy.” Liu Peichao said. To address this, the company is studying data回流 (data circulation), hoping that the 100,000 arms shipped out each year can return data from different scenarios. In terms of model types, last year VLA (vision-language action models) plus reinforcement learning could already be deployed, but the endgame is world models. However, the amount of data needed to build world models requires tens of millions to even hundreds of millions, so it will need 3–5 years to plan for.

Unitree management said the company is also advancing verification of the VLA technology roadmap and laying out world models. In 2026, the company’s R&D budget is expected to grow to 700 million yuan, with key focus on embodied large models, world models, and product iterations.

Huge amounts of information, precise interpretation—All in the Sina Finance APP