Humanoid Robot Learning Patents: Who Leads, Where Gaps Are 2026
- Filing only started moving in 2022 and jumped to 27 families in 2025 — this is a field still in its first real filing wave, not a mature one.
- China dominates the receiving offices with 24 filings against 15 from the United States and a single PCT entry, so venue strategy already looks regional rather than global.
- Manipulator claims (B25J) cover 36 of 40 families while control-systems (G05B) and vision (G06T) sit far behind — the mechanical/control split is lopsided, not the learning stack itself.
A young, fast-accelerating filing field
Patent activity in humanoid robot learning and control is concentrated in a narrow window: filings were essentially flat through 2017-2021, then began climbing in 2022 and reached a peak of 27 families in 2025. That timing lines up with the broader push toward whole-body control and sim-to-real transfer methods entering commercial humanoid programmes. The 40 total families in this dataset make it a small, still-forming corpus rather than a saturated one.
Publication lags filing by roughly 18 months, so the 2026 count of 4 understates what has actually been filed this year. The technology composition skews heavily toward B25J manipulator and robot-arm claims, with G06N (AI models) and G05B (control systems) as secondary clusters — meaning the learning algorithms themselves are less densely claimed than the mechanical and control-system implementations they run on.
Let an AI agent run this analysis on your own technology
Pick a task. Every answer cites the patents behind it.
Filing trend and technology composition
Two views of the same 40-family dataset: how filing volume has moved year over year, and where those filings sit across the IPC classification system.
Filings accelerated from a standing start
Zero filings in 2017, a single filing at the 2022 midpoint, then a run to 27 in 2025 before the partial 2026 count of 4. The shape is a late, steep acceleration rather than steady growth — most of the corpus is only two to three years old.
B25J manipulator claims dominate the classification spread
36 of 40 families touch B25J (manipulators and robots), far ahead of G06N (10, AI models), B62D and G05B (8 each, vehicle/steering and control systems), and smaller counts in G06F, G06T, A61B and E21B. The imbalance suggests claim drafting has centred on the physical robot and its control loop more than on the learning model architecture itself.
Shares are the percentage of the 40 records in scope. A patent can carry several IPC classes, so the shares add up to more than 100%.
Go deeper on Humanoid Robot Learning and Control with Eureka
This page is one run against one query. Ask Eureka your own question about humanoid robot learning and control and every answer comes back with the patent numbers behind it.
Try EurekaThe most-cited filings so far
Multi-motion switching control method and system for humanoid robot based on imitation learning
Filed by Shandong University, this application performs imitation learning via a Generative Adversarial Network and dynamically adjusts the sampling probability for each motion skill based on how well the humanoid robot executes it. The stated aim is uniform mastery across different motion skills, letting the robot combine skills flexibly while mitigating mode collapse — a known failure mode in GAN-based imitation learning.Published 2026-06-04, illustrating how recent this corpus is even at the representative-filing level.
| # | Publication no. | Patent title | Citations |
|---|---|---|---|
| 1 | CN118664586A | 一种结合周期奖励的人形机器人步态模仿学习方法 | 8 |
| 2 | CN121061903A | 一种基于强化学习的人形机器人双臂具身操作方法 | 7 |
| 3 | US20260070221A1 | Bipedal action model for humanoid robot | 6 |
| 4 | CN117961888A | 一种基于强化学习控制的人形机器人物体抓取方法及装置 | 6 |
| 5 | US20260097492A1 | Annotation model for humanoid robot data | 5 |
| 6 | CN119427324A | 一种力反馈外骨骼及人形机器人模仿学习方法 | 4 |
| 7 | US20220324109A1 | Method and apparatus for controlling multi-legged robot, and storage medium | 4 |
| 8 | CN120116218A | 人形机器人模仿学习方法、装置、计算机设备及存储介质 | 3 |
| 9 | CN119427360A | 基于模仿学习的人形机器人多运动切换控制方法及系统 | 3 |
| 10 | CN119501927A | 面向非结构化环境的轮腿式机器人运动控制方法、系统、装置、存储介质及计算机设备 | 2 |
Citation counts here reflect an early, small corpus and favour the oldest available records; they signal early influence, not necessarily current commercial relevance.
Patent titles are shown in the language they were filed in, not translated, so that each record stays verifiable against the original filing — a translated title will not match in Eureka or in any national register. Each row carries its publication number; clicking a row searches Eureka by that number.
Put your own technology through the same analysis
Eureka on the web
When you want the answer in the next five minutes.
The agent works the prompt against patents and technical literature, citing every source.
Run your analysis now →MCP server & REST API
When it has to run inside your own pipeline.
Patent search, landscape analysis and assignee resolution as MCP tools. Drop them into any agent framework, or call REST directly.
Browse MCP servers →What the numbers mean for filing strategy
Three read-throughs from the trend, geography and classification data that matter for anyone deciding where to file next or where to expect resistance.
The claim space is still open, but closing fast
A field that goes from zero to 27 annual families in under a decade has not yet settled its dominant claim language. Filing now still lets an applicant shape how core sim-to-real and imitation-learning terms get defined in this corpus, but that window narrows every year the growth rate holds.
Filing venue is regional, not global yet
With only a single PCT filing recorded, most applicants are choosing a home jurisdiction rather than pursuing broad international protection. That leaves freedom-to-operate gaps for competitors filing directly into whichever of the two major offices a given applicant skipped.
Mechanical and control claims outweigh model claims
G06N (AI models) appears in only 10 of 40 families versus 36 for B25J. Learning algorithms are frequently claimed as a method step within a broader robot-control system rather than as a standalone model architecture, which changes how a freedom-to-operate search should be scoped.
Eureka can read the same corpus for gaps instead of for coverage: under-claimed branches adjacent to humanoid robot learning and control, with the prior art for and against each one.
| Assignee | Co-assignee | Shared families |
|---|---|---|
| Luming Technology (Suzhou) Co., Ltd. | Luming Robotics Technology (Shenzhen) Co., Ltd. | 1 |
Only one co-assignee pairing appears in the dataset, between two affiliated Luming entities, indicating most work here is filed by single organisations rather than joint ventures or research partnerships.
Who is filing, and where momentum is shifting
Recent-year momentum is mixed: several assignees that were active a year or two ago show zero or sharply declining filings in the latest period, while newer single-digit filers are just entering the dataset.
Early movers are pulling back
Several assignees active in prior years, including Digital Artificial Intelligence Corp and Beijing Songyan Dynamics, show sharp year-over-year declines into the latest period rather than sustained growth, which is common in a corpus this young where filing bursts are followed by pauses.
New academic entrants are still arriving
Liaoning University shows a single filing in the latest year with no prior baseline, consistent with a field where universities are still establishing initial positions rather than building sustained portfolios.
Some established filers show no recent activity
Harbin Institute of Technology and South China University of Technology both show zero filings in the latest year, with the latter down 100% year over year, suggesting research output has not yet converted into fresh filings or has shifted to other classifications.
| Assignee | Recent year | YoY |
|---|---|---|
| Digital Artificial Intelligence Corp | 1 | -89% |
| Beijing Songyan Dynamics Technology Group Co., Ltd. | 1 | -50% |
| Liaoning University | 1 | — |
| Shandong University | 0 | -100% |
| Harbin Institute of Technology | 0 | — |
| South China University of Technology | 0 | -100% |
| Shanghai Kepler Robotics Co., Ltd. | 0 | -100% |
| Luming Technology (Suzhou) Co., Ltd. | 0 | -100% |
Where to take this analysis
The trend and classification data point to specific next steps depending on whether the goal is freedom-to-operate, portfolio building, or competitive tracking.
Run a freedom-to-operate check before filing
With B25J claims covering 36 of 40 families, a new filing on robot-arm or whole-body control methods needs a targeted search against that subclass specifically, not just the G06N AI-model literature.
Search prior art in EurekaTrack assignee momentum quarterly
Several assignees show double-digit percentage swings year over year in a corpus this small; a single new filing can flip momentum, so tracking needs to be more frequent than an annual review.
Set up assignee monitoring in EurekaScope claims around under-claimed branches
Vision-conditioned control and sim-to-real transfer for bipedal locomotion remain thin relative to core manipulator claims, leaving room for narrowly drafted first claims in those areas.
Explore white space in EurekaCommon questions about humanoid robot learning patents
This dataset covers 40 published patent families filed between 2015 and mid-2026 under a search combining humanoid, bipedal and legged robot terms with reinforcement learning, imitation learning, sim-to-real transfer and whole-body control learning language. Filing was essentially flat through 2021, then accelerated sharply, peaking at 27 families in 2025. Because publication lags filing by around 18 months, the 2026 figure is understated and the true total for that year will be higher once more filings publish.
China leads by receiving-office count with 24 filings, compared with 15 for the United States and a single filing routed through the WIPO PCT system. The near-absence of PCT filings suggests most applicants are protecting a single home jurisdiction rather than pursuing broad multi-country coverage, which leaves gaps for competitors to file directly in whichever major office an applicant has not covered.
Sim-to-real transfer refers to methods that train a robot's control policy in simulation and then adapt it to work on physical hardware, addressing the gap between simulated and real-world dynamics. It is one of the core search terms behind this dataset and sits at the intersection of the G06N (AI models) and G05B (control systems) classifications. Because it appears less densely claimed than manipulator hardware (B25J), it remains a technically specific area where narrowly scoped claims can still be filed.
The dataset shows a mix of established filers and new entrants rather than a single dominant leader. Some assignees active in earlier years, such as Beijing Songyan Dynamics and Digital Artificial Intelligence Corp, show declining filings in the most recent year, while others like Liaoning University appear only in the latest period with a single filing. This pattern is typical of a young field where portfolio leadership has not yet consolidated.
US20260151908A1, assigned to Shandong University, claims a multi-motion switching control method using imitation learning via a Generative Adversarial Network, with a sampling probability that adjusts dynamically based on how well the robot performs each motion skill. It specifically addresses mode collapse in GAN-based imitation learning by encouraging uniform skill mastery. This blocks approaches that use the same dynamic-sampling mechanism for multi-skill GAN imitation learning, but alternative sampling strategies, non-GAN imitation methods, or different mode-collapse mitigation techniques would sit outside its specific claim language.
Research Humanoid Robot Learning and Control in depth with Eureka
Go past this page: query the whole humanoid robot learning and control corpus yourself, in your own scope.
Every answer comes back with patent numbers you can open.
Disclaimer. This page is generated from Patsnap Eureka data drawn from a limited snapshot of global patent and scientific-literature records, and is provided for general information and reference only.
Patent data carries inherent limitations: recent filings (typically the most recent 18–24 months) are under-counted due to standard publication lag; counts may be reported at either a patent-family or a patent-record basis and are not always directly comparable; classification, applicant-name, and citation data may contain errors, duplicates, or omissions; and the underlying search query defines and constrains the scope shown. As a result, the analysis may be incomplete or inaccurate and may not reflect the full technology landscape.
Nothing on this page constitutes an exhaustive prior-art, novelty, freedom-to-operate, or validity search, nor does it constitute legal, financial, investment, or professional advice, and it should not be relied upon as such. Any patent, commercial, or strategic decision should be verified independently and reviewed with qualified patent, legal, and domain professionals. Patsnap makes no warranties, express or implied, as to the accuracy, completeness, or fitness for any particular purpose of the information presented.
Machine translation. Assignee and organisation names originally recorded in Chinese, Japanese or Korean have been rendered into English by an AI translation step so that the tables stay readable. These renderings are best-effort and may not match a company’s registered English name; the original name is what the underlying patent record carries, and it is what any Eureka query launched from this page uses.