Embodied AI Direction Study: Starting From Our 'Reservoir Computing Moat,' Finding a 3-Year Direction That Both Earns Money and Survives

⚠️ Read the market figures as 'orders of magnitude,' not precise values. The market sizes / growth rates cited here come from third-party research firms such as Grand View, MarketsandMarkets, and Fortune Business Insights, and for the same market, different firms routinely differ by 2–10x (especially for buzzwords like "Physical AI / humanoid"). Every figure is sourced, and the conclusions rely only on judgments like "this is a tens-of-billions-of-dollars scale, growing 20%+ annually" — they never bet on any precise value. This piece builds on the moat conclusions from the prior article, Existing Results Summary.

🌐 What This Piece Answers + Research Method

The real question for launching this venture is not "is Embodied AI hot or not," but: a 10-person team, standing on the company's RC edge-AI moat and starting from our existing 4 demos — over the next 3 years, where can we go to both earn money and stay alive?

Methodologically, we deliberately work backwards: rather than entering through whole-machine hype like "which robot dog to buy / whether to build a humanoid" (that's a capital black hole, see below), we start from the moat we already hold and go out to the market to find the opening that 'fits this blade best and where someone is already paying.' Each direction is judged on three hard criteria: ① is there a real paying party ② can the capital/team scale support it ③ can it generate cash within 3 years.

---

📊 First Cut: A Bubble Check on the "Embodied AI / Physical AI" Mega-Trend

Start with the big numbers (mark the order of magnitude, don't trust the precision):

Bubble-check conclusions (3 points):

1. Whole machine / humanoid = capital black hole; a 10-person team should not touch the body itself. The humanoid "trillion-dollar TAM" is a speculative extrapolation: Goldman estimates $38B by 2035²³; Citi shouts $7T by 2050, Morgan Stanley shouts $5T²⁵— the same thing differs by roughly 180x across investment banks, and that dispersion itself shows it's a story, not a business. Hard evidence: even Tesla, on its early-2026 earnings call, admitted that the few hundred Optimus units it has deployed are "primarily for learning rather than production, still in the R&D stage"; the de-hype voice Rodney Brooks bluntly says deployable dexterous manipulation for humanoids is "clumsy until after 2036"²⁴. For the next 3 years, humanoids are essentially pre-revenue pilots; a 10-person team that bets on the body will die.
2. What is actually flowing money 'now' is the 'software & module layer' that makes machines reliable / perceptive — edge AI, anomaly detection, safety perception. This layer scales up commercially in exactly 2026. And this money is backed by a real "installed base": per IFR data, the global operational stock of industrial robots has reached 4.66M units²¹— these several million robots, plus tens of thousands of cobots/AMRs, are all a paying installed base that "needs anomaly detection / health monitoring / safety modules."
3. Japan's national policy is the biggest tailwind for this (and its very center overlaps perfectly with our 3 years). From FY2026, ¥1 trillion (about $6.3B) in 5-year AI support²⁸; most critically, NEDO's "Generative-AI Foundation Model & Data Platform for Robots" program, scale ¥20.5B / FY2025–2029 (recipient AIRoA, with participation from the Matsuo Lab at the University of Tokyo)¹⁶— this is the very center of Japan's embodied-AI national policy, and its time window overlaps exactly with our venture's horizon. At the same time, METI's robot-friendly (robofure) policy locks in four sectors — facility management / food / retail / logistics warehouses — for environmental retrofitting¹⁷. "Robots filling the jobs no one is willing to do" is Japan's most certain demand for the next 3 years¹².

---

🇯🇵 Insert: Japan Demand "Confidence Grading" — Don't Get Misled by Buzzwords; Pick the Highest-Certainty Scenarios

Japan's robot demand is driven by demographics + the coercive force of regulation, not hype. But certainty varies a lot between scenarios; when launching, aim for the "★★★" ones and stay conservative on the "⚠️ overhyped" ones. This table directly determines where the main-line vertical should anchor:

How to read it: Main line A (equipment anomaly detection) naturally fits the three ★★★ — "logistics equipment / infrastructure inspection / manufacturing production lines"; eldercare, though often treated as the hot frontier, has measured adoption that is extremely low, so treat it as long-term exploration and keep it out of the 3-year cash plan.

---

🇯🇵 Insert 2: Japan's Fiscal Support — What It Funds / How Much / Who Can Get It

The bubble-check mentioned "¥1 trillion in 5-year AI support" and "NEDO's ¥20.5B foundation-model & data-PF." The chairman will ask: what exactly does this money fund, how much (subsidy rate / cap), and can a 10-person team like IVIS actually get it? This section splits Japan's money into three types — first, one rule to remember: commissioned-cost ≠ subsidy ≠ grant; whether you can "directly apply and get paid" depends entirely on which type it is.

¥1 Trillion / 5-Year AI Support: a composite package, not one fund you can "claim"

• The top-level basis is the "AI Basic Plan" (Cabinet decision of 2025-12-23, under the AI Promotion Act)²⁹. But ⚠️ the "¥1 trillion" figure is NOT written in the plan text itself — it is an FY2026 budget / press figure; the plan lists only policy directions, while amounts run through the budget process³⁰.
• It is not a single fund, but a composite package: domestic foundation models, data-center and other infrastructure, Physical-AI R&D and demonstration, startup support, and SME AI-adoption subsidies in parallel. The plan's named targets include "Physical-AI industrial pioneering-adoption support [METI]," "support for startups with innovative AI tech," "promotion of SME AI adoption (subsidy)," and "AI-robot public-demand creation + R&D demonstration"²⁹.
• → How IVIS should read it: this ¥1 trillion is not a window where you "claim one slice directly"; the real ammunition entrance is the specific NEDO solicitations below.

NEDO × AIRoA ¥20.5B: a "commissioned project," not an openly applicable subsidy

• The commissionee is the AI Robot Association (AIRoA), program scale ¥20.5B / 48 months (FY2025–2029)¹⁶.
• ⚠️ Its nature is a "commissioned project" from NEDO to the AIRoA consortium — not a subsidy companies apply for openly — advanced in three stages (collect/curate high-quality on-site robot data → general foundation model → use-case-specific models + social implementation). Participants / sub-commissionees include Waseda, Toyota, AIST, and the Matsuo·Iwasawa Lab at the University of Tokyo³¹.
• → A 10-person external team like IVIS can only participate indirectly: ① join the AIRoA consortium / become a sub-commissionee, ② contribute data·components to the data PF (feed the data), ③ become a downstream user of the foundation model / PF. A company cannot "apply for a subsidy" from this ¥20.5B on its own (it's a commission scheme).

The NEDO solicitations IVIS Can "Directly Apply" For (target / cap / subsidy rate)

• Deep-Tech Startup Support Fund (DTSU) — fund total ¥87.7B, program period FY2023–2032¹⁰; its legal form is a grant (subsidy rate < 100%), targeting deep-tech startups needing "long horizons + large R&D capital":
  • Three phases — STS (commercialization, early) / PCA (commercialization, late) / DMP (mass-production demonstration) — per-phase cap roughly ¥0.3–2.5B, total up to ¥3B / max 6 years, with stage-gate continuous support³²;
  • Subsidy rate 2/3 or 1/2: if within the prescribed period you raise external funding equal to ≥1/6 of the eligible cost, you get 2/3; otherwise 1/2 (you co-fund the rest)³². ⚠️ Per-phase caps vary by solicitation round, so go by the requirements of the actual round before applying.
• "Multimodal Foundation Model Development Program Targeting AI Robots / Physical AI" solicitation — targets include companies / universities, and its nature is a composite of "research (commission · joint research · subsidy)", solicitation window 2026/03–04⁹; ⚠️ the specific budget / subsidy rate is not stated in the HTML, so mark it "unconfirmed" and check the solicitation guidelines before applying.

The Three Types of Money, Split (one-line distinction)

One-cut mnemonic: Commissioned cost = doing the state's work, fully funded, outputs lean to the state; subsidy / grant = doing your own work, partial, you co-fund, outputs are yours.

Operational conclusion for IVIS: the only indirect participation is NEDO × AIRoA ¥20.5B (commissioned, not separately applicable) — via consortium participation or feeding the data PF; the prime target you can directly apply and get paid is DTSU grant (subsidy rate 2/3, per-phase ¥0.3–2.5B, total up to ¥3B / 6 years), and IVIS's existing NEDO commission track record makes a follow-on application more credible.

---

🎯 Second Cut: Aiming the Moat at the Market Layers

The Embodied AI tech stack, roughly from bottom to top, is: body mechanics → actuation/power → sensing/perception → state estimation/anomaly detection → planning/control → safety → cloud/digital twin. A 10-person software team should not touch the bottom two layers (capital-heavy) or the top cloud layer (the giants' home turf); what we should eat is the middle three layers: perception — anomaly detection — safety — lightweight control — and this is exactly where our 4 demos have already proven out.

One-line positioning: define the team as a "supplier of 'anomaly detection / safety perception / lightweight edge control' software & modules for Embodied AI," not a "robot company." The moat (RC: ultra-lightweight, low-power, online incremental learning, strong at time-series anomaly detection) sits squarely on this layer — capital-light, licensable, contract-able, and able to grow into a product.

Two 'profitable adjacent service layers' — not reliant on the RC moat, but suitable to eat together with the team / hardware partner, to pad Year-1 cash:

• System integration (SI) = the most realistic money layer for a small team. The industry is highly fragmented with no winner-takes-all; it's project-based / service-type revenue with almost zero product-R&D capital, and the exit path is clear too (JR Automation was acquired by Hitachi for $1.43B²⁰. Collaboration with "the other company's hardware factory manager" naturally lands here — pick one vertical scenario (food packaging / e-commerce sortation / loading-unloading for a specific process step) and go deep.
• Functional-safety compliance consulting = a regulatory must-have, capital-light. ISO 10218-2:2025 has now formally folded in ISO/TS 15066, making the integrator legally the "manufacturer" of the whole workcell, who must produce a risk assessment and CE marking¹⁸; the new standard's rollout will keep generating compliance demand.
  • ⚠️ Cautionary tale (must read): the star company that made an "intelligent safety-guarding product," Veo Robotics, raised a cumulative $56M and obtained TÜV certification, yet in the end was sold off cheap to Symbotic for just $8.7M (about 1/6 of the funding raised)¹⁹. Lesson: building a safety 'product' is a cash-burn pit; doing safety 'consulting/integration' is the money a small team can actually make.

---

🔍 Third Cut: Rapid Assessment of 5 Candidate Directions (With Scores)

Scoring dimensions: clarity of paying party / moat fit / capital lightness / 3-year cash / competitive intensity (🟢 good 🟡 medium 🔴 poor).

Direction A　Equipment Predictive Maintenance · Time-Series Anomaly Detection (🟢🟢🟢🟢🟡) — The Steadiest Near-Term Cash Cow

• Market: predictive maintenance 2025/26 ≈ $14–17B → 2034 ≈ $97B, CAGR ~24–28%⁴; within it, vibration monitoring holds ~40% share, and acoustic monitoring has the highest growth (CAGR 42.7%)⁵— and anomaly detection on vibration/acoustic/current time-series is exactly the home turf of our Demo ①'s "train on normal data → alert on deviation" approach.
• Paying party: factory/equipment operators, already paying for this (one hour of downtime is a huge loss). Demand is mature; no market education needed.
• Moat fit: extremely high. RC's "edge online incremental learning" directly solves this market's biggest pain point — every machine's operating conditions differ, cannot be labeled in advance, and require fast on-site adaptation (the likes of Augury rely on cloud-side large models + their own high-end sensors, expensive and heavy).
• Competition: intense but layered. The leaders — Augury (raised $369M, the most), Senseye (Siemens), IBM Maximo, GE APM, Tractian, Nanoprecise (acoustic)⁶— focus on "high-value large equipment + cloud platforms." The differentiation opening left for us = the "low-cost, pure-edge, no-cloud, on-site self-learning" mid/small equipment / privacy-sensitive / offline operating conditions.
• Verdict: this is the main-line cash-cow candidate. Start with a "vibration/current/acoustic edge anomaly-detection box or software module," bootstrap via contract work + NEDO, then productize.

Direction B　Robot Safety Module · Sensorless Collision Detection / Robot Health Monitoring (🟢🟢🟢🟡🟡) — Closest to Demo ①, Licensable

• Market logic: collaborative robots (cobots) / mobile robots are exploding, and safety is a hard requirement, regulation-driven. The industry clearly distinguishes "collision detection vs. collision avoidance," and model-based 'sensorless collision detection' (monitoring the deviation between expected dynamics and measured) is a well-recognized route — "real-time response, no added hardware"¹⁵. This is exactly what Demo ① does; our RC version is one such implementation.
• Paying party + installed base: robot OEMs / cobot makers / AMR makers — they buy the module that "makes my machine safer/more reliable," cheaper than building it themselves. The pool is backed by IFR data: the global installed base of industrial robots is 4.66M units²¹, cobots make up about 10% of annual installations, plus millions of AMRs (Amazon's own use alone has passed 1M units) — all a potential install surface for health-monitoring / safety modules.
• Moat fit: high, directly reusing Demo ① (ROS2 output, already engineered).
• Risks: ① OEMs may build it in-house (sensorless collision detection is already a known method) → we must differentiate via RC's lightness + online adaptation (no retraining across different machine models); ② long sales cycle (selling to OEMs means getting into their product line).
• Verdict: side-line / module-licensing candidate. It fits naturally with collaboration with "the other company's hardware factory manager" — he provides the channel and whole machine, we provide the safety/health-monitoring software layer.

Direction C　Sound × Anomaly · Edge Health/Security/Footstep Recognition (🟡🟢🟢🟡🟢) — Blue Ocean but Needs to Find Buyers

• Basis: Demo ④ already names the applications — voice-based physical-condition anomaly detection, footstep-sound burglary prevention, pure-CPU online learning.
• Market: acoustic monitoring is the fastest-growing sub-item within predictive maintenance (CAGR 42.7%, as above); security/care are hard-requirement labor-shortage scenarios in Japan¹¹.
• Risks: the paying party is less clear than "factory equipment maintenance," demand needs education; care/health touches medical compliance.
• Verdict: an exploratory side line, usable as a NEDO project or a joint PoC with a care/security SI, not as main-line cash.

Direction D　Edge Adaptive Control (close to Demo ②) (🔴🟢🔴🔴🟡) — Most Cutting-Edge, Most Distant; Don't Live On It

• Basis: Demo ②'s "RC world model + reinforcement learning" for sample-efficient edge control. Technically sexy, high paper/demo value.
• Risks: farthest from a paying product; control-safety certification has a high bar, and the customer-trust cycle is long.
• Verdict: keep it as an 'R&D banner / NEDO project / recruiting calling card', not expecting it to monetize within 3 years.

Direction E　Riding the Tailwind of Japan's "Physical AI Domestic Foundation Model" National Policy (🟡🟢🟢🟡🟡) — Grab Funding + Stake Out an Ecosystem Niche

• Basis: the very center of Japan's embodied-AI national policy is NEDO's "Generative-AI Foundation Model & Data Platform for Robots" program (¥20.5B / 2025–2029, recipient AIRoA)¹⁶, plus the "Multimodal Foundation Model Development Program Targeting AI Robots / Physical AI"⁹. The foundation model itself is out of reach for a 10-person team, but "edge-side lightweight anomaly detection / safety components + data collection" is a missing piece for landing the foundation model, which can be the entry as a supporting project / part of the data PF — and the time window (2025–2029) overlaps perfectly with our 3-year roadmap.
• Leverage: we already have a NEDO commission track record (Demo ④'s RC additional-learning result is itself a NEDO-commissioned deliverable), so a follow-on application carries high credibility.
• Verdict: a cross-cutting funding/endorsement source, not a standalone product direction — A/B/C should all be packaged into NEDO applications as much as possible.

Assessment summary: main line = A (predictive maintenance / edge anomaly detection); module side line = B (robot safety, with the hardware partner); exploratory side line = C (sound × anomaly); banner = D; funding ecosystem = E running through all of them.

---

⚔️ Fourth Cut: Competitive Reality Check — Where Exactly Are We Differentiated

Honestly, "edge anomaly detection" is not an empty field; we must answer "why us" head-on:

• vs. cloud predictive-maintenance giants (Augury/Senseye): they are expensive, heavy, require the cloud, and require their own high-end sensors. Our opening is "cheap, pure-edge, usable offline, on-site self-learning" — serving the mid/small equipment and privacy/offline conditions they look down on.
• vs. mainstream TinyML (isolation forest / TCN / autoencoder on MCU): TinyML can already train in 1.2–6.4 s on a microcontroller, detect in <16 ms, and run anomaly detection in 80 KB RAM¹⁴— this shows "edge anomaly detection" is feasible and has competitors; the moat is not "can it run on the edge" but whether RC's "online incremental learning + time-series dynamic patterns" is superior in scenarios of "continuously drifting operating conditions where you must learn while in use." This point must be proven with our own PoC data, not just told as a story.
• vs. the RC hardware camp (TDK's analog reservoir-computing chip⁷, Quantum Computing's photonic RC NeuraWave⁸: they make RC chips/hardware; we make the RC 'algorithm + application + online-learning engineering + ROS2/production-line integration' software layer — we can in fact become an application user / partner of their chips (TDK is in Japan, a potential partner rather than a pure competitor). This shows RC is moving from academia toward commercialization, and the time window is in our favor.

One-line differentiation: don't say "we know RC," say "we can turn RC into a 'time-series anomaly-detection / safety module that fits into small on-site devices and keeps learning while in use even offline,' and integrate it into robots and production lines." This is an engineering + application moat, not a mere algorithm gimmick.

---

🧭 Fifth Cut: Business Model & 3-Year Roadmap (Make the Chairman Dare to Invest, Make the Team Able to Survive)

Core rhythm: cash first → productize one vertical → grow recurring.

10-person staffing suggestion (drawn from the company's AI and advanced-technology departments):

• 2–3 RC/ML algorithm (the moat, people already in place)
• 2 edge/embedded engineering (FPGA/MCU/Pi, practiced in Demos ③④)
• 1–2 robotics/ROS2 + control integration (Demos ①②)
• 1 data/signal processing (vibration/acoustic/current features)
• 1 solutions/pre-sales + 1 PM/business (interface with the hardware partner's channel, write NEDO applications, negotiate customers) — these two are the easiest to overlook, yet they are the key to 'staying alive.'

Division of labor with "the other company's hardware factory manager": he provides the hardware body + channel + on-the-ground execution; we provide the software/perception/anomaly-detection/safety layer — a textbook "independent software-layer player" play, avoiding the capital-heavy body and using exactly our cards.

---

⚠️ Sixth Cut: Failure Modes and Avoidance

When small embodied / edge-AI teams die within 3 years, they nearly always die on one of these:

1. Going to fight over the whole machine body → crushed by capital. Avoidance: only ever do the software/module layer; rely on the partner for hardware.
2. Falling into the 'custom-project trap' → every customer is rebuilt from scratch, a product never grows, and per-capita output is locked. Avoidance: in Year 1, deliberately pick contract work that 'can converge into the same product'; the reuse rate of the Nth project must rise.
3. Tech self-indulgence with no paying party → living on Direction D. Avoidance: the main line must pick predictive maintenance where 'the customer is already paying,' not the sexiest control.
4. Sales cycle longer than the cash runway → slow contracting when selling to OEMs/large firms. Avoidance: use NEDO grants + contract work to pad cash and outlast the long sales cycle; position the chairman's investment as 'runway supplement,' not 'waiting for returns.'
5. A moat that's a story with no data → replaced by TinyML. Avoidance: do a comparison PoC as early as possible, use real operating-condition-drift data to prove RC online-learning's advantage, and write it up as a reproducible benchmark.

---

✅ Recommended Conclusion: One Main Line + a Set of Bets

Main line (bet 70% of resources): Direction A — equipment predictive maintenance / edge time-series anomaly detection. Rationale: the clearest paying party, the best moat fit, the lightest capital, the most likely to generate cash in 3 years; triple tailwind of Japan's labor shortage + manufacturing fundamentals + government national policy.

Side bets (bet 30%, hedge + position):

• B (robot safety / health-monitoring module) — feed it directly into the collaboration with the "hardware factory manager" as a licensable module;
• C (sound × anomaly) — as an exploratory NEDO/PoC, not expected to be cash that year;
• D (edge adaptive control) — kept as an R&D banner and recruiting calling card;
• E (NEDO Physical AI national policy) — a funding and endorsement source running through all directions; package A/B/C into applications as much as possible.

One-line version for the chairman: "We don't build robot bodies (that's a capital black hole); we build 'the edge-AI brain that makes robots and factory equipment reliable' — standing on the company's existing NEDO-track-record Reservoir Computing moat, we first stay alive on contract work and national funding, and within 3 years turn 'equipment anomaly detection' into a replicable product, earning the most certain money of the labor-shortage era."

---

🔎 To Be Verified & Intelligence Gaps (Please Correct With Real Data Before Deciding)

• Large definitional gaps in market figures: all the market sizes/growth rates above come from third-party research firms, and the same market differs by several times across sources; use them only for order-of-magnitude judgment, never write them into an investment-return model as precise values.
• The moat's 'hard evidence' is still missing: RC online learning's advantage over mainstream TinyML (isolation forest/TCN/autoencoder) in 'operating-condition drift' scenarios is currently an inference; we need our own comparison benchmark before we can claim it externally (this is a Year-1 must-fill).
• Willingness to pay and pricing not validated: the real customer quotes, procurement cycles, and decision chains for Directions A/B are not yet researched — we need the hardware partner's channel to run 3–5 real customer interviews.
• The specific NEDO solicitation and grant amounts: this piece cites the existence and caps of NEDO funds / Physical-AI solicitations¹⁰; which specific round, whether they can be stacked, and IVIS eligibility need confirmation with a NEDO contact.
• Competitors' actual penetration locally in Japan: the actual share and pricing of Augury/Senseye/Tractian in Japan's small/medium manufacturing has not been verified one by one.