Positioning and terminology
This article provides industry judgments based on this edition's 10 frontier news items (as of 2026-09-08). Conclusions aim to be specific, verifiable, and schedulable. All judgments are annotatedBasisandConfidence(High / Medium-High / Medium / Medium-Low), with data sources and timestamps noted; unverifiable points are explicitly marked asSpeculation.
I. Competitive Landscape: A Three-Layer Structure Is Taking Shape
Layer 1: Embodied Hardware and Complete-System Manufacturing (Already Highly Concentrated)
In H1 2026, global humanoid robot shipments were approximately 19,100 units, +272% year over year, with Chinese manufacturers accounting for over 97%; AgiBot shipped approximately 8,400 units (44%) and Unitree approximately 5,900 units (31%), together nearly 75%; Unitree listed on the STAR Market on August 19 (issue price CNY 15.80 per share), and AgiBot has initiated the Hong Kong listing process. (Source: third-party institutions and public reports, 2026-08)
Assessment: Concentration in the hardware layer is increasing faster than expected; the Matthew effect has already taken hold.This means that embodiment manufacturers will become the largest buyers of data procurement, and at the same time will extend upstream by building their own capture capabilities.(The AGIBOT-incubated Mifeng Technology is one example).
Confidence: High.
Layer 2: Models and Algorithms (Crowded but Not Yet Decided)
VLA architectures are converging, but absolute capability remains low — in RoboDojo's first leaderboard, top simulation models averaged an 8.80% success rate, with the best real-robot result at 12.8%, compared with 76.03% and 100% for humans respectively (2026-07). Players are dense both domestically and internationally (Physical Intelligence, NVIDIA, Tencent, UBTech, Xinghaizhou, Xiaomi, Ant Group, etc.).
Assessment: No winner-take-all outcome will emerge in the model layer over the next 12 months, but there will beDivergence of Capability Dimensions(memory, long-horizon, fine manipulation). For data providers, the more crowded the model layer, the better — we do not need to bet on who wins, only to serve everyone.
Confidence: Medium–High.
Layer 3: Data Supply Side (Being Repriced by Capital and Industry)
Data is shifting from "laboratory output" to "supply-chain capacity": crowdsourced collection settled by daily wages has already emerged domestically; manufacturers estimate annual capacity at tens of millions of hours based on "10,000 devices × 4 hours/day × 250 working days"; State Grid plans to invest 6.8 billion yuan in 2026 to procure 8,500 embodied-AI devices, with rapid growth in procurement by local training centers and practical-training bases.
Assessment:The data supply side is the most undervalued link in the current industry chain, and at the same time the one with the shortest entry window.. The window is roughly 12–24 months—once a given device becomes the default hardware for crowdsourced capture, the switching cost is no longer just the hardware, but the sum of formats, toolchains, and model-side conventions.
Confidence: Medium–High(The window duration isSpeculative judgment, depending on the ramp-up of embodied hardware and the speed of model iteration).
II. Industry Chain Opportunities: Four Entry Points
| Opportunity | Description | Our Relative Advantages | Barriers to Entry | Recommended posture |
|---|---|---|---|---|
| Collection Hardware | Head-mounted/wrist multi-camera, wearable | Existing products and mass-production experience, synchronization and calibration capabilities | Medium | Hold firm and strengthen |
| Data Capacity Services | Deliver quality-inspected data by scenario | Integrated Hardware + Workflow Capability | Medium–High | Prioritize Investment |
| Formats and Toolchains | Conversion tools, data-card specifications, quality-inspection tools | Neutrality (no in-house model development) | Low–Medium | Open Source for Ecosystem |
| Compliance and Quality Certification | De-identification, authorization, traceability, quality-inspection reports | Controllable at the collection end, complete chain | Low | Productize as a premium offering |
Explicitly Out of Scope: General embodied foundation models (competing with customers), embodied hardware and dexterous-hand manufacturing (misaligned with positioning), general simulation platforms (ecosystem already mature).
III. Partnership and Investment Risks
Partnership Opportunities (by Priority)
- 1.Establish joint "data–evaluation" validation with model companies: Use the language of public leaderboards (RoboDojo's five dimensions, RxBrain-Bench) to demonstrate how our data improves specific capability dimensions. The persuasiveness of such joint experiments far exceeds scale-based marketing.Highest priority.
- 2.Coordinate real2sim interfaces with simulation platform providers: Convert real captured data directly into interactive environments to amplify the value of each individual data point.
- 3.Jointly define time reference and paired-capture specifications with tactile sensor manufacturers: pre-emptively claim the niche of a cross-sensor unified latent space.
- 4.Maintain alignment with standards and evaluation bodies: Aligning quality dimension definitions in advance is equivalent to securing acceptance criteria authority early.
Investment risks (stop-loss conditions must be defined)
| Risk | Description | Trigger Signals | Response |
|---|---|---|---|
| In-house capture by robot manufacturers | Customers extend upstream and bypass us | Leading customers begin developing and mass-producing their own head-mounted devices | Shift to a "device + quality inspection + compliance" service portfolio, moving value from hardware to process |
| Data price war | Crowdsourced oversupply causes unit-price collapse | Public quotation quarter-over-quarter decline >30% | Immediately pivot to high-value verticals (medical/laboratory, precision assembly) and hedge through differentiation |
| Format lock-in | A model company dominates the format and creates exclusivity | Mainstream open-source projects support only a single format | Align across multiple formats in advance and open-source conversion tools to maintain neutrality |
| Compliance incident | Improper face/privacy handling leads to data recall | Industry-wide penalty cases emerge | Deploy compliance packages upfront, establishing authorization chains and de-identification audits |
| Sudden technical route shift | World-model approaches reduce the importance of visual data | Visual share in mainstream model training mixes drops significantly | Maintain modality neutrality and simultaneously lay out tactile/force and multimodal time references |
IV. Action Recommendations for the Next 12 Months (Executable, Scheduled, with Metrics)
Q4 2026 (Oct–Dec): Lay the Foundation
| Initiatives | Deliverables | Validation metrics |
|---|---|---|
| Open-source format conversion tools (bidirectional LeRobot / MCAP / RLDS) | GitHub repository + documentation | Cited by ≥3 third-party projects; customer integration time reduced by ≥50% |
| Compliance package productization | License agreement templates, de-identification workflows, batch traceability records | First-pass legal approval rate from major enterprise customers ≥80% |
| Free sample dataset + complete data card | 3 scenario samples + format documentation | Downloads ≥500; conversion rate into commercial pipeline within 30 days ≥15% |
| Geometric quality inspection pipeline goes live | Quality inspection report for each data batch | Report fields cover reprojection error, pose drift, and synchronization deviation |
Q1 2027 (Jan–Mar): Build Differentiation
| Initiatives | Deliverables | Validation metrics |
|---|---|---|
| Real-time quality guidance at capture (coverage score + synchronization status indicators) | Firmware + edge applications | Reduction of ≥30% in the number of episodes required to reach the same policy success rate |
| Long-horizon and Failure Sample Acquisition Protocol | Protocol documentation + annotation specifications | Completeness rate of 100% for all three sample types (success / failure / correction) in delivered data |
| Conduct joint validation with 1–2 model companies | Joint experiment report | Provide reproducible improvement margins on the memory/long-horizon dimension |
Q2 2027 (April–June): Mode change
| Initiatives | Deliverables | Validation metrics |
|---|---|---|
| Shift from "selling devices" to "selling effective hours" | Quality-inspected data packages delivered by scenario | Data services revenue share ≥25% |
| Three vertical-scenario data packages (industrial assembly / logistics picking / laboratory manipulation) | Scenario-specific datasets | Contract value per scenario and repurchase rate |
| Online calibration and drift monitoring tools | Tools delivered with the device | Reduction in calibration maintenance labor hours for customer fleet deployments ≥50% |
Q3 2027 (Jul–Sep): Scale-up
| Initiatives | Deliverables | Validation metrics |
|---|---|---|
| Crowdsourced data collection network pilot | Collection network + settlement and quality-control workflows | Valid hours per month, cost per hour |
| Standardized integration of tactile/force channels | Hardware interface specification + data card fields | Supports ≥3 mainstream tactile sensors |
| Customer Success (model-side) team | Walk clients through the full "data → training → evaluation" pipeline | Customer model success rate improvement, NPS |
V. Key Judgments at a Glance (with confidence levels)
| Assessment | Confidence | Basis | Consequences if the judgment is wrong |
|---|---|---|---|
| The data supply side is the most undervalued link in the industry chain | Medium–High | Robot unit volume up 272%, data capacity becomes the bottleneck, crowdsourcing takes shape | Medium: can fall back to a pure hardware model |
| Window of 12–24 months | Medium (Speculation) | Hardware replacement cost rises as ecosystem lock-in increases | Medium: miss the ecosystem niche, but not fatal |
| Not developing a large model in-house is the right choice | High | Compute cost structure + neutrality issues regarding competition with customers | Low |
| Compliance capability will shift from a cost item to a premium item within 12 months | Medium–High | Legal review is front-loaded in large-enterprise procurement processes | Low–Medium |
| The model layer will not become winner-take-all | Medium–High | Absolute capability remains low; capability dimensions are diverging | Low: we are neutral by default |
| Probability of a price war emerging in 2027 is not low | Medium | Crowdsourced supply expands rapidly | High: vertical-domain differentiation must be planned in advance |
VI. One-Sentence Strategy
In the cycle where data becomes the core production factor of embodied AI, what TOBI should occupy is not "a position in some model" nor "a position in some hardware," but "the measurement and delivery standard for effective data capacity"—so the industry collects with our devices, delivers in our format, and accepts against our metrics.
Disclaimer and maintenance: All data cited in this article are labeled with source and time; capacity estimates, window length, price-war probability, etc. areSpeculative judgmentand should be re-reviewed as the market changes. We recommend reviewing risk trigger signals monthly.