The Robotic Ecosystem Map
A Landscape Analysis of Embodied Physical AI from Agriculture to Humanoids
A two-edition ecosystem map of embodied physical AI, comparing agricultural robotics in 2025 with humanoid robotics in 2026 to identify where ownership consolidates, where markets remain contested, and where the durable governance gaps sit above the machine.
Surendra Reddy27 min read
Abstract
Embodied physical artificial intelligence has moved from demonstration into commercial deployment across agriculture, warehouse logistics, manufacturing, and human-proximate settings within a narrow window, and the capital committed to the category now substantially exceeds the independently verified evidence of value at scale. This paper is a landscape analysis of that category, structured as an ecosystem map rather than a product survey. It decomposes an embodied machine into a layered stack, characterizes each layer by its business contestedness rather than only its engineering function, and identifies the structural gaps in the landscape that are candidates for research, investment, or venture formation. The analysis proceeds in two editions. The first, grounded in agricultural robotics as of early 2025, decomposes a field robot into six layers and surfaces four gaps: the agronomic prescription layer, an open-source protocol bridge, an edge-compute transition, and a financing gap between venture-scale capital and machine-scale unit economics. The second edition, grounded in the humanoid robotics landscape as of mid-2026, extends the decomposition to a seventh layer, an interaction and embodiment layer that a task-specific machine does not require and a human-proximate machine cannot function without, and adds a foundation-model layer that the agricultural edition never needed. Across both editions the paper's central finding is stable: durable value in embodied physical AI does not concentrate in any layer of the machine but in a governance layer that sits above the stack, the prescription or defensible-action layer, which remains structurally unclaimed by any current player in every domain examined. The paper documents the platform and startup topology of both editions with dated, source-graded market evidence, and states its assumptions and prognosis explicitly so that a later reader can check whether the ground has shifted.
The form factor changed completely. The shape of what remains ungoverned did not.
This Musewoods Journal edition preserves the analysis, evidentiary qualifications, and references of the August 2026 SSRN working paper while giving the argument the Journal's long-form editorial treatment.
1. Introduction
This paper is the empirical companion to a specification paper that sets out the Zai composable architecture for embodied physical AI. Where the specification paper defines the architecture and its properties formally, this paper supplies the landscape analysis that motivates and tests it: who is building in the category, along which strategic lines, and where the structural gaps in the technology and business topology actually sit. The two papers are designed to be read together but each stands alone, and this paper's findings do not depend on the architecture specification for their validity, only for their theoretical framing.
The method of this paper is ecosystem mapping. Rather than surveying products or ranking companies, it decomposes the embodied machine into a layered stack and asks of each layer three questions: who owns it today, how contested is it as a business, and is there a structural gap at or above it worth building around. This method treats the category as an economic and architectural system rather than a collection of machines, and it is designed to let a founder, investor, or research organization stop treating embodied physical AI as one undifferentiated technology problem and instead reason layer by layer about where durable value and open opportunity reside.
The paper is organized as two editions of the same map. The first edition, summarized in Section 2, analyzes agricultural robotics as of March 2025 and is documented in full in the companion architecture paper and prior work. The second edition, which forms the substance of this paper from Section 3 onward, extends the map to humanoid robotics as of mid-2026. The two-edition structure is deliberate. Holding the analytical method constant across a fifteen-month interval and a complete change of form factor is what lets the paper distinguish findings that are artifacts of one domain from findings that are structural features of embodied physical AI as such.
1.1 Contribution
This paper makes three contributions. First, it provides a dated, source-graded landscape analysis of the humanoid robotics platform and startup topology as of mid-2026, organized by strategic posture rather than by product category. Second, it extends a layered ecosystem map from six layers to seven, documenting which layers are invariant to form factor, which transform, and which is genuinely new, and characterizing each by business contestedness. Third, it identifies and grades the structural gaps in the landscape, distinguishing gaps that are competitive from gaps that are structural, and argues that the largest and most durable gap, the governance layer above the stack, is stable across both editions and both form factors.
1.2 A Note on Method and Evidence
The market evidence in this paper is dated and graded. Funding totals, valuations, and deployment counts for privately held companies are self-reported by those companies or their investors and are treated as company-disclosed figures rather than independently audited totals, a distinction the paper marks explicitly wherever such figures appear. Claims about a company's strategic posture are inferred from its stated positioning and disclosed actions, not from private information. Where a single technical result is load-bearing for the analysis, the paper cites the primary source directly. The paper's contestedness characterizations are analytical judgments derived from this evidence, and the paper states its assumptions explicitly in Section 8 so that those judgments can be checked against future developments.
2. The First Edition in Brief: Agriculture as of March 2025
The first edition of this map decomposed an agricultural robot into six layers, data and orchestration, connectivity, positioning and navigation, edge compute, on-robot software, and actuation and sensing, and surfaced four gaps. The first was the agronomic prescription gap, the absence of any formal, accountable layer governing who decides what treatment gets applied where and by what authority. The second was a narrow protocol gap, the absence of a mature open-source bridge between the standard robot middleware and the tractor-to-implement communication protocol that remains essential in agriculture. The third was an edge-compute transition, an anticipated shift in on-machine compute sourcing. The fourth was a financing gap, the mismatch between venture-scale capital expectations and the long, seasonal, service-heavy path to unit-level profitability that agricultural machines actually follow.
The first edition's central finding was that agricultural robotics was constrained less by machine capability than by ecosystem architecture, and that durable value was migrating away from single machines toward the layers that govern trusted, accountable field action. Market developments through mid-2026 sharpened rather than contradicted that finding. Among the agricultural companies the first edition surveyed, the most capital-intensive full-stack machine builder collapsed and had its autonomy technology absorbed by a heavy-equipment incumbent, another wound down and was acquired by a customer, while the companies pursuing more disciplined, less vertically integrated strategies continued to grow. The pattern the first edition identified, that attempting to own every layer of the machine is a capital trap, held under a year of additional evidence.
3. The Platform Imperative, Restated for a Different Body
The first edition opened with a claim about agriculture that generalizes further than expected: the diversity of real-world tasks defeats any monolithic, single-purpose machine, which means the durable architecture has to be modular, built from horizontal rails that commoditize quickly and specialized components that compete on top of them. Humanoid robotics arrives at the same architectural instinct from the opposite direction. Rather than building one machine per task, the category is betting that one general-purpose body, paired with a foundation model that learns new tasks from demonstration rather than hand-coded programming, can absorb the diversity that previously required many specialized machines.
That bet is now well capitalized and partially tested in the field rather than purely theoretical. Figure AI closed a Series C exceeding one billion dollars in September 2025 at a thirty-nine billion dollar post-money valuation, with NVIDIA, Brookfield, Salesforce, and Microsoft among the participants, and by early 2026 its manufacturing facility was producing humanoids at a cadence of roughly one unit every ninety minutes. Agility Robotics deployed its bipedal Digit platform across nine customer facilities with more than sixty-five thousand cumulative operating hours logged as of its June 2026 announcement of a public listing through a two and a half billion dollar merger.
Tesla's Optimus program moved past a one-thousand-unit threshold inside Tesla's own factories by early 2026, though the company stated on its January 2026 earnings call that those units remained primarily for learning and data collection rather than verified productive work, a distinction this paper treats as evidentially important rather than incidental.
The platform imperative that organized the first edition still holds. What has changed is which layer of the stack the imperative is being fought over, and the remainder of this paper works through that question layer by layer.
4. Platform Topology: Three Bets on What a Body Is For
The first edition found three agricultural equipment manufacturers making three different structural bets about which layer of the stack to own. The humanoid category in 2026 shows a comparable pattern of divergent bets, though the bets concern what kind of body and what kind of intelligence will scale rather than which layer to control. The topology matters for the same reason it did in the first edition: it locates the real competitive lines before a layer-by-layer analysis can be drawn honestly.
| Platform posture | Core bet | What would validate the bet |
|---|---|---|
| Vertical integration | A generalized body and brain will outperform task specialization once the model generalizes far enough. | Repeatable commercial deployment across tasks and customers. |
| Narrow proven deployment | Revenue, operating hours, and task reliability matter more than speculative generality. | Expansion of paid deployments without sacrificing unit economics or reliability. |
| Manufacturing scale | Automotive-style production scale becomes the decisive cost and distribution moat. | High-volume production paired with verified productive work rather than learning-only deployment. |
4.1 Figure AI: Vertical Integration of Body and Brain
Figure's position is the clearest of the three. After ending a collaboration with OpenAI in early 2025, the company's stated reasoning was that solving embodied AI at scale requires vertically integrating robot intelligence rather than depending on a foundation model from a partner whose roadmap the robotics company cannot control. Its Helix system is a single neural network driving both navigation and manipulation, extended to whole-body control in a January 2026 release, and its data strategy uses human video at scale to bootstrap capability without robot-by-robot manual training.
That vertical integration is capital-intensive, and the company's own production figures illustrate the cost of the strategy. Figure increased output from roughly one unit per day to one per hour over several months in 2026, producing more than three hundred fifty third-generation units with reported first-pass yields above eighty percent. The commercial proof point most often cited, a multi-month deployment at an automotive plant handling tens of thousands of vehicles, is real but bounded, and the same automaker's decision to use a competitor's platform for its first European humanoid pilot is a credible signal that an early lead with one customer has not generalized into preferred status even with that customer.
This paper reads Figure as the clearest instance of the generalization bet: that a vertically integrated body and brain will out-compete specialized machines once the underlying model generalizes well enough.
4.2 Agility Robotics: Narrow, Proven Deployment Over Generality
Agility's posture inverts Figure's. Rather than betting on a generalized body that learns any task, it has prioritized task-specific deployment of its Digit platform inside warehouse and logistics workflows, and its stated differentiation is that Digit generates revenue from paying commercial customers rather than research pilots. Its nine-facility footprint, disclosed multi-year orders exceeding three hundred million dollars, and announced public listing are evidence of a company choosing proven narrowness over speculative breadth.
In the vocabulary of the first edition, this is closely analogous to the agricultural retrofit-and-deployment posture: a strategy that trades the upside of generality for the credibility of revenue and operating hours that competitors betting on generality cannot yet show. Agility's investor base reads as a bet on that credibility rather than on a speculative foundation-model breakthrough.
4.3 Tesla: Manufacturing Scale as the Moat
Tesla's Optimus program occupies a third position, betting that the decisive advantage belongs to whoever can manufacture humanoids at automotive volumes first, with capability following the path electric-vehicle production followed: a difficult, expensive ramp that becomes a durable cost advantage. The company frames this comparison explicitly, likening the program's production start to its earlier mass-market vehicle ramp.
The evidentiary picture here requires more caution than Figure's or Agility's, because the program's public statements carry a documented history of missed timelines across several prior years. As of the company's January 2026 earnings call, the roughly one thousand units inside its factories were characterized as primarily for learning and data collection rather than verified productive work. The manufacturing-scale bet is therefore currently a capital and infrastructure commitment, evidenced by large announced capital expenditure and the conversion of existing vehicle production lines, rather than a demonstrated operational one, and this paper weights it accordingly.
4.4 What the Pattern Reveals
Three bets, generalized body, narrow proven deployment, and manufacturing scale, and none has yet produced the unambiguous commercial proof that the agricultural companies of the first edition had already produced by their own 2025 date. Those companies had real customers paying for measurable outcomes when the first edition was written. The humanoid category, at this edition's date, has substantial capital, real production ramps, and a small number of real deployments, but not the proof of value at scale that would let an observer distinguish the winning bet from the others.
That asymmetry, more capital and attention chasing less proven value than the category the first edition surveyed, is the single most important finding of this section, and it conditions the layer-by-layer analysis that follows.
5. The Foundation Model Layer: New Ground the First Edition Never Mapped
The first edition had no equivalent of this section, because agricultural robots in 2025 ran narrow perception models trained for one task rather than general-purpose models meant to transfer across tasks. Humanoid robotics has produced the latter, and it changes the substance of the edge-compute and on-robot-software layers rather than merely adding a vendor to an existing list.
Three providers anchor this layer as of this edition's date. One pairs a vision-language-model reasoning backbone with a motor-control module and reports substantial task-success gains from synthetic training data. A second adds three-dimensional spatial perception and on-device operation, allowing the model to run locally rather than depending on a cloud round-trip for each decision. A third has demonstrated the most closely watched generalization results in the category, including performance on tasks the model was never explicitly trained on, achieved by synthesizing fragments of related experience with general pretraining.
The most consequential single result for this layer's trajectory is the EgoScale study, which trained a vision-language-action policy on more than twenty thousand hours of action-labeled egocentric human video and reported a log-linear scaling law between human-data scale and validation loss, with that loss correlating with downstream real-robot performance (Zheng et al., 2026). This is the first strong evidence that robotics foundation models follow the same data-driven scaling dynamics that characterized large language models, a reading independently echoed in the 2026 robotics theses of major technology investors.
If the finding holds under further scrutiny and replication, it implies that the edge-compute and on-robot-software layers are converging toward the compute-and-data-scaling dynamics that reshaped language models, which would further centralize the position of whichever providers command the most pretraining data and compute. This paper treats the EgoScale result as strong but recent, and its scaling-law implication as a working assumption rather than a settled fact, consistent with the broader observation that scaling laws are empirical regularities rather than guarantees.
6. The Startup and Platform Layer, Extended
The first edition's startup layer surfaced agricultural companies differentiated by function. The humanoid equivalent is differentiated less by function, since most of the category is still pursuing general-purpose capability rather than a narrow task, and more by go-to-market posture and geography. One affordability-focused entrant has moved beyond demonstration into intralogistics pilots at automotive plants while targeting a sub-fifty-thousand-dollar price point. Europe's most heavily funded humanoid company closed approximately one and a fifth billion dollars in early 2026 to build what it terms cognitive robots for industrial and daily settings.
The scale of activity in China exceeds anything the first edition's agricultural survey approached. One manufacturer reported producing its ten-thousandth humanoid by late March 2026, scaling from roughly one thousand units the prior year, and lower-cost platforms starting near sixteen thousand dollars have become a credible price-disruption threat to Western platforms priced an order of magnitude higher. Industry analysis places China's share of global robotics component supply chains at roughly seventy percent as of 2025, a concentration that mirrors the vertically consolidated supply-chain advantage established earlier in electric vehicles, and that several analysts argue Western makers will struggle to counter without deliberate, multi-year supply-chain investment.
One deployment outside the factory-and-warehouse pattern is analytically important because it evidences the distribution argument central to this paper more directly than any factory pilot. A major airline, responding to a working-age population projected to decline by roughly a third over the coming decades, deployed lower-cost humanoid platforms for baggage handling, container transport, and cabin cleaning at a major international airport in 2026. The stated rationale was explicit: the airport was built for people rather than for wheeled machines, so the humanoid form factor was chosen to fit infrastructure that already exists rather than infrastructure that would require rebuilding. This is the distribution argument, that a human form factor inherits a human-built environment as its operating context, evidenced in a live commercial deployment rather than argued from first principles.
7. The Seven-Layer Map: Ownership and Contestedness
The first edition's central artifact was a table assigning ownership and a contestedness rating to each of six layers. This edition extends that table to seven, and the extension is substantive rather than cosmetic, because grounding each row in the evidence above changes several of the original contestedness ratings, not only the row count.
| Layer | Function | Ownership as of this edition | Contestedness |
|---|---|---|---|
| Actuation and sensing | Physical components touching the world, now spanning many modalities in one body | Fragmented component vendors; no dominant humanoid-specific actuator supplier, with China holding roughly 70 percent of component supply chains | Low at the component level; high geopolitical stakes at the supply-chain level |
| On-robot software | Middleware and real-time control coordinating many actuation modalities | Open-standard middleware remains the substrate; no proprietary humanoid-specific owner has emerged | Low; the most settled layer, as in the first edition |
| Edge compute and foundation models | On-robot perception, reasoning, and motor control | A small number of foundation-model providers; increasingly the layer platforms build on rather than around | High; absent from the first edition and now the most actively contested |
| Positioning and navigation | Indoor, human-density spatial awareness rather than open-field positioning | No dominant proprietary owner; each platform builds its own stack | Medium; harder than the agricultural case, not yet consolidated |
| Connectivity | Machine-to-cloud data flow | Standard wireless infrastructure; unchanged from the first edition | Low; commodity infrastructure |
| Data and orchestration | Fleet dispatch, logging, and system of record across heterogeneous tasks | Fragmented across each platform vendor's own cloud stack | Medium and rising as fleets scale past pilot size |
| Interaction and embodiment | Behavior around and with people: proximity, social legibility, consent, and context-switching | No current player; genuinely unowned | Unrated; no market has yet formed to contest it |
Above all seven layers sits the prescription layer, also termed defensible action, which governs the authority to decide and defend a consequential action. It remains outside and above the stack exactly as it sat above the original six, unclaimed by any current player.
The prescription layer is not another machine layer. It is the authority boundary around the machine.
8. Reading the Map: Which Layers Hold, Which Transform, Which Is New
Five of the original six layers extend into the humanoid case with the scope of the work changing but not the function. Actuation and sensing extends rather than breaks: the engineering bar rises, evidenced by rapid production ramps and substantial payload gains across platform generations, without changing what the layer is fundamentally for. On-robot software holds its open character, because coordinating more actuators requires a more complex graph of processes rather than a different middleware philosophy. Connectivity and the data-and-orchestration plane hold their original shape, with only the texture of what they coordinate changing from one repeated task to many heterogeneous ones across a shift.
Edge compute is the layer that has genuinely transformed rather than merely extended. The first edition treated it as a hardware-sourcing question, a choice among ruggedized on-board computers. The humanoid edition shows the substantive contest has moved up a level of abstraction, from which processor to which foundation model, and the incumbent edge-compute provider's position has strengthened rather than weakened in the transition, because the leading embodied models run on the same hardware lineage the first edition already flagged as ascendant. Positioning and navigation extends to a harder version of the same problem, finer-grained and shared with moving people rather than open and empty, but remains recognizably the same function.
The interaction and embodiment layer is the one genuine structural addition, and the airport deployment is the clearest evidence that its absence is being felt in practice rather than only in theory. A humanoid working alongside human ground crew at a major airport operates in continuous, contextual proximity to people that a field robot never encounters, and no platform surveyed in this edition has published a framework for governing that proximity as a distinct, auditable concern separate from task execution. This layer is vacant where machines operate away from people and demanding where they operate among them, which is precisely why it is the layer that distinguishes the humanoid case from the agricultural one.
9. The Gaps This Edition Surfaces
Three gaps stand out, building on the four the first edition surfaced rather than replacing them.
| Gap | Type | Why it matters |
|---|---|---|
| Interaction and embodiment | Structural absence | Human-proximate machines need governable behavior around people, yet no distinct auditable layer is established. |
| Foundation models | Concentration | The layer exists, but data and compute dynamics may compound advantage toward a small number of providers. |
| Capital versus proven value | Evidentiary | Investment and valuation are moving faster than independently verified commercial proof. |
| Prescription / defensible action | Structural absence above the stack | No current machine-layer owner governs the authority, evidence, conditions, and accountability of consequential action. |
The interaction and embodiment layer is the largest new gap, and it is structural rather than competitive in the same sense the prescription gap was structural in the first edition. No platform surveyed here has published a framework for governing proximity and social context as a distinct, auditable layer separate from task execution, which means whoever builds it first is not competing against an incumbent's head start but filling a genuine void.
The foundation model layer is the second gap, and it is a gap of concentration rather than absence. A small number of credible providers anchor it, and the scaling-law evidence suggests the layer will compound in favor of whoever commands the most pretraining data and compute rather than whoever holds the cleverest narrow model. A platform builder that builds on one of these providers rather than attempting to compete with them is making the same buy-rather-than-build choice the first edition recommended for agriculture's commodity layers.
The third gap is evidentiary rather than architectural, and it deserves to be named as plainly as the financing gap the first edition surfaced. The capital flowing into humanoid robotics is running well ahead of independently verified commercial proof. One platform's valuation reportedly exceeds a major investment bank's projection for the entire global humanoid market a decade out, against modest projected near-term revenue, and the most manufacturing-committed program's own leadership has stated its deployed units are not yet performing verified productive work.
The gap between capital deployed and value proven is itself the binding constraint on the category's near-term trajectory, in the same structural sense the financing gap constrained agricultural robotics in the first edition.
10. Lineage: What Carried Forward and What Changed
This edition kept the analytical spine of the first deliberately intact rather than starting fresh. The platform imperative, the case for modular, horizontal-rail architecture over monolithic single-purpose machines, carried forward unchanged and was simply tested against a different actuator. The equipment-manufacturer topology of the first edition became a platform topology in this one, three structural bets rather than three manufacturer strategies, but the analytical move of asking what each player believes about where value settles is identical. The startup layer extended rather than replaced, adding a foundation-model layer the first edition never needed because no agricultural robot in 2025 ran a model meant to transfer across tasks.
What did not carry forward unchanged is the contestedness picture. Edge compute moved from a hardware-sourcing contest to a foundation-model contest. Positioning and navigation became harder rather than staying flat. And the prescription layer, unclaimed above six layers in the first edition, now sits unclaimed above seven, with the new interaction and embodiment layer enlarging the unclaimed territory rather than reducing it. That is the single most important continuity between editions: the form factor changed completely, and the shape of what remains ungoverned did not.
11. Assumptions and Prognosis
Three assumptions carry the weight of this edition's reading, stated explicitly so a later reader can check whether the ground has shifted. The first is that the foundation-model layer continues consolidating around its small set of credible providers rather than fragmenting, an assumption qualified by the recency of the scaling-law evidence and its lack of independent replication at the scale language-model scaling laws have reached. The second is that the gap between capital and proven value persists rather than resolving quickly, which implies that the relative positions of the leading platforms a year out will be determined more by which converts pilots into repeatable revenue than by which raises the next largest round.
The third is that no regulator issues binding guidance on human-proximate humanoid operation within the next twelve months, since if one does, the unowned status of the interaction and embodiment layer converts from an opportunity into a compliance deadline.
Given those assumptions, the prognosis follows the shape the first edition's did. The bottom layers, actuation, on-robot software, and connectivity, continue commoditizing. The foundation-model layer consolidates further, making it a buy-rather-than-build decision for any new entrant exactly as edge compute became in agriculture.
The interaction and embodiment layer either gets built by a credible new entrant within this window or stays open, because nothing in the platform topology gives the leading platforms a structural incentive to solve it before a regulator forces the question. And the prescription layer remains where it has been across both editions: the most valuable gap in the system, unclaimed by anyone, and unlikely to be claimed quickly regardless of which body eventually carries the intelligence that tries.
12. Conclusion
Across a fifteen-month interval and a complete change of form factor, from a task-specific field robot to a general-purpose humanoid, the analytical method of this map produced a stable central finding. Durable value in embodied physical AI does not concentrate in any layer of the machine, all of which either commoditize or consolidate toward a small number of providers, but in a governance layer that sits above the machine entirely and remains unclaimed in every domain examined.
The humanoid form factor did not change this finding. It enlarged the ungoverned territory, adding an interaction and embodiment layer that a machine operating among people cannot function without and that no current player owns. The category's binding near-term constraint is not machine capability, which is advancing quickly, but the gap between the capital committed to it and the value independently proven, and its most durable long-term opportunity is the governance layer that neither edition of this map found anyone building.
Durable value does not settle in the machine layer that moves fastest. It settles where consequential action becomes governable, attributable, and defensible.
Appendix A: Layer Comparison Across Editions
The following table summarizes how each layer's contestedness changed between the first edition (agriculture, 2025) and this edition (humanoids, 2026).
| Layer | First edition (2025) | This edition (2026) |
|---|---|---|
| Actuation and sensing | Low; fragmented components | Low at component level; supply-chain concentration is now a strategic concern |
| On-robot software | Low; open standard settled | Low; unchanged |
| Edge compute and foundation models | Medium; a hardware-sourcing question | High; a foundation-model contest |
| Positioning and navigation | Medium; open-field positioning | Medium; harder indoor, human-dense variant |
| Connectivity | Low; commodity | Low; unchanged |
| Data and orchestration | Medium and rising | Medium and rising |
| Interaction and embodiment | Not present | New; unrated, unowned |
| Prescription / defensible action | Above the stack; unclaimed | Above the stack; unclaimed |
A Note on AI-Assisted Research and Intelligence Provenance
This paper was developed with the assistance of two AI systems under active research and development at 451 Labs. Astra, the Adaptive Strategic Thesis Reasoning Agent, supported the market signal research and monitoring that underpins the platform and startup analysis in this paper, tracking relevant signals across the agricultural and humanoid robotics landscapes. Clearstory, the knowledge and intelligence provenance system and the primary realization of the Astra thesis, supported the tracking of source lineage and claim provenance throughout, helping maintain the dated, source-graded evidentiary discipline this paper holds itself to. Neither system is a shipped product. Both are theses under active development at 451 Labs, and their contribution to this paper is itself part of the validation process this research is designed to run. All analytical judgments, contestedness ratings, gap identifications, and prognosis claims remain the author's own.
References
Reddy, S. (2025). Zai: The composable robot. An architectural blueprint for a scalable agricultural robotics platform [Working paper]. 451 Ventures.
Reddy, S. (2026). Zai: A composable layered architecture for embodied physical AI across agriculture, industry, and care [Companion working paper]. 451 Labs / 451 Ventures.
Reddy, S. (2026). Regenerative Prosperity Model 2.0 [Working paper]. SSRN. https://ssrn.com/abstract=7162098
Reddy, S. (2026). The RPM Decisioning System: Five Regeneration Gates, two operating disciplines, and the stewardship loop [Working paper]. SSRN. https://ssrn.com/abstract=7244703
Zheng, R., Niu, D., Xie, Y., Wang, J., Xu, M., Jiang, Y., Castañeda, F., Hu, F., Tan, Y. L., Fu, L., Darrell, T., Huang, F., Zhu, Y., Xu, D., & Fan, L. (2026). EgoScale: Scaling dexterous manipulation with diverse egocentric human data (arXiv:2602.16710). arXiv. https://arxiv.org/abs/2602.16710
Churchill Capital Corp XI. (2026, June). Agility Robotics to go public through $2.5 billion merger with Churchill Capital Corp XI [Form 8-K, Exhibit 99.1]. U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/0002074973/000121390026071287/ea029548401ex99-1.htm
Figure AI. (2025, September). Figure exceeds $1B in Series C funding at $39B post-money valuation [Press release]. https://www.figure.ai/news/series-c
GeekWire. (2026, June). 'Digit' maker Agility Robotics to go public in $2.5B deal. https://www.geekwire.com/2026/digit-maker-agility-robotics-to-go-public-in-2-5b-deal-heres-what-the-filings-say-about-its-finances/
Origin of Bots. (2026, April 21). Agility Robotics' Digit moves beyond pilots at Amazon, Toyota, and GXO. https://www.originofbots.com/news/agility-robotics-digit-moves-beyond-pilots-now-handling-real-warehouse-work-at-amazon-toyota-and-gxo
The AI Insider. (2026, May 1). Figure AI ramps up production to one humanoid robot per hour. https://theaiinsider.tech/2026/05/01/figure-ai-ramps-up-production-to-one-humanoid-robot-per-hour/
The Robot Report. (2025, September 19). Figure AI passes $1B with Series C funding toward humanoid robot development. https://www.therobotreport.com/figure-ai-raises-1b-in-series-c-funding-toward-humanoid-robot-development/
Bessemer Venture Partners. (2026, April). Bessemer predicts: Robotics and physical AI. https://www.bvp.com/atlas/bessemer-predicts-robotics-and-physical-ai
KraneShares. (2026, May 12). Humanoid robotics in 2026: The race from pilot to platform. https://kraneshares.com/humanoid-robotics-in-2026-the-race-from-pilot-to-platform/
Note on sourcing. Funding totals, valuations, and unit-deployment counts for privately held companies are self-reported by those companies, their investors, or pending regulatory filings, and are treated throughout as company-disclosed figures rather than independently audited totals. The manufacturing-scale claims of the most production-committed humanoid program are treated with additional caution given that program's documented history of missed prior timelines, and the prognosis weights that bet as unproven rather than demonstrated. The EgoScale result is a peer-community preprint whose scaling-law implication, while independently echoed by major technology investors, has not yet been replicated at the scale language-model scaling laws have reached, and is treated here as a strong but working assumption. Category-level market-size projections referenced in background research are forecasts dependent on category definitions and were not relied upon for any specific company claim.