Robotics foundation-model layer evidence packet v1

Date: 2026-06-12 Owner: Hugo / Genius Team Agent: Finance / Charlie AGT-002 Status: public-safe research artifact Visibility: PUBLIC, no trade recommendation

0. One-line answer

The robotics model/data layer has moved from pure demo narrative toward source-backed S3/S4 technical-platform evidence: Physical Intelligence shows open and progressively more general VLA-style robot policies, while Skild AI shows an omni-bodied foundation-model strategy with large financing and industrial deployment partnerships; however, neither public record yet proves S5 scaled commercial economics.

1. Core question

If humanoid / embodied robotics scales, does the AI model layer become a horizontal value layer like an operating system, or does value stay mostly inside each robot OEM's vertical stack?

Current answer as of 2026-06-12:

The model layer is now investable as a research lens because multiple primary sources show cross-embodiment / cross-task model ambition, open-source tooling, industrial partnerships, and data-flywheel strategies. 🟢
The model layer is not yet proven as a standalone commercial value-capture layer because public sources still lack recurring revenue, license pricing, deployed-seat/robot counts, gross margin, retention, or customer ROI/payback. 🟢 absence in reviewed sources / 🟠 inference

2. Market-definition lens

This artifact covers the AI / software / data layer, not the robot-body OEM layer and not component supply chain.

Boundaries:

Included: robot foundation models, VLA / policy models, cross-embodiment control, data flywheel, simulation, human-video learning, deployment partnerships where the model is the claimed differentiator.
Excluded: hardware supplier revenue, robot ASP, reducer/actuator BOM, public-equity trade ideas.

Do not double-count: A robot OEM using a model stack is not automatically evidence that the model provider captures economics. We need contract economics, usage pricing, attach rate, or revenue disclosure before upgrading to S5.

3. Evidence table

Company / layer	Evidence as of date	What is proven	What is not proven	Source grade	Signal
Physical Intelligence / π0	2024-10-31	π0 is described by the company as a general-purpose robot foundation model trained on broad robot data and VLM pretraining; it spans images, text, and actions and outputs low-level motor commands. The article says π0 used Open X Embodiment, internet-scale pretraining, and PI datasets from 8 distinct robots; it starts from a 3B-parameter VLM and can output motor commands up to 50 times/sec.	No disclosed customer contract, revenue, runtime in paid deployment, or gross margin.	🟢 company research post	S3 technical-platform evidence
Physical Intelligence / openpi	2025-02-04	PI released code and weights for π0 through the experimental openpi repository; stated release includes pretrained π0, fine-tuned checkpoints, inference code for real/sim robot platforms, and fine-tuning code. PI says 1-20 hours of platform data was sufficient in its own experiments for fine-tuning to several tasks, with caveat “your mileage may vary.”	Open-source availability does not prove enterprise adoption, revenue, reliability, or customer ROI.	🟢 company research post / GitHub reference	S3 ecosystem signal
Physical Intelligence / π0.5	2025-04-22	π0.5 is described as a VLA with open-world generalization; PI says it can deploy out-of-the-box in new homes and generalize to new settings such as cleaning a kitchen/bedroom not seen in training. The article explicitly says the current model is “far from perfect.”	No scaled in-home deployment, no uptime/intervention-rate disclosure, no customer economics.	🟢 company research post	S3/S4 technical generalization signal
Physical Intelligence / π0.7	2026-04-16	π0.7 is described as a steerable general-purpose model with a step-change in generalization; PI says it can perform a wide range of dexterous tasks with similar performance to fine-tuned specialists and follow new language commands / tasks not seen in training data.	Company-reported experiments are not the same as scaled commercial deployment; no revenue or customer ROI.	🟢 company research post	S4 model-capability signal, still pre-commercial
Skild AI / company thesis	Website accessed 2026-06-12	Skild states it is building a unified, omni-bodied brain to control any robot for any task; listed applications include security/inspection, mobile manipulation, and autonomous packing.	Website positioning does not prove model performance, revenue, or deployment economics.	🟢 company website	S3 positioning signal
Skild AI / Series A	2024-07-09	Skild says it raised US$300m Series A at US$1.5bn valuation; stated use is scaling the model and team.	Financing does not prove commercial traction or durable value capture.	🟢 company announcement	S3 capital signal
Skild AI / Series C	2026-01-14	Skild says it raised US$1.4bn Series C led by SoftBank, valuing the company at over US$14bn, with participation from NVentures/NVIDIA, Macquarie Capital entities, Bezos Expeditions, and others.	Valuation and capital do not prove product-market fit or unit economics.	🟢 company announcement	S4 capital/strategic-backers signal
Skild AI / data strategy	2026-01-12 and 2025-07-29	Skild argues teleoperation alone cannot reach foundation-model scale; claims it uses internet video, large-scale simulation, and targeted real-world data. In the Jan 2026 post, Skild says it can fine-tune a model to new skills by watching videos alone plus <1 hour of robot data.	This is company-claimed technical evidence, not audited deployment economics.	🟢 company research posts	S3/S4 data-bottleneck signal
Skild AI / industrial partnerships	2026-03-19	Skild says it is partnering with ABB Robotics, Universal Robots, and NVIDIA to deploy the Skild Brain across industries and applications without task-by-task reprogramming.	Partnership language does not disclose deployed robot count, contract value, production usage, uptime, or ROI.	🟢 company announcement	S4 industrial-channel signal
Skild AI / Zebra robotics arm	2026-04-15	Skild says it acquired Zebra Technologies' robotics division, formerly Fetch Robotics, to deploy its omni-bodied brain across warehouses and add to its data flywheel; the post references Zebra's Symmetry Fulfillment orchestration platform and real-time data from Zebra wearable devices.	Acquisition and platform assets do not prove the model has replaced classical workflows at scale or created software gross margin.	🟢 company announcement	S4 go-to-market/data asset signal

4. Interpretation: why this matters

The model layer matters because the main bottleneck in general-purpose robotics is not only hardware cost; it is robust generalization under messy real-world conditions.

Source-backed catalysts:

Cross-embodiment ambition is no longer only conceptual: PI's π0 used data from 8 distinct robots and Skild claims a model spanning quadrupeds, humanoids, tabletop arms, and mobile manipulators. 🟢 as of 2024-10-31 / 2026-01-14
Data strategy is becoming a central competitive variable: PI emphasizes robot data + internet-scale VLM pretraining; Skild emphasizes simulation + internet video + targeted real-world data and explicitly attacks the teleoperation bottleneck. 🟢 as of 2024-10-31 / 2025-07-29 / 2026-01-12
Open-source / ecosystem moves may lower experimentation cost: PI released openpi code and weights on 2025-02-04, including fine-tuning and inference code. 🟢
Industrial channel validation is emerging: Skild's 2026 posts cite partnerships with ABB Robotics, Universal Robots, NVIDIA, and acquisition of Zebra/Fetch robotics assets. 🟢
Capital allocation has stepped up: Skild moved from US$300m Series A at US$1.5bn valuation in 2024 to US$1.4bn Series C at over US$14bn valuation in 2026. 🟢

5. Counter-evidence / bear case

Public evidence still has three large gaps:

Commercial gap: no reviewed public source discloses recurring software revenue, license price, take rate, attach rate, gross margin, or ARR for PI or Skild robotics models. 🟢 absence in reviewed official sources / 🟠 interpretation
Deployment gap: partnership and acquisition announcements do not disclose robot count, paid utilization, intervention rate, uptime, repeat orders, or customer ROI/payback. 🟢
Value-capture gap: if Tesla, Figure, Unitree, Agility, UBTECH, or other OEMs internalize model stacks, horizontal model providers may create technical value without capturing most economics. 🟠 scenario, needs evidence

6. Stage classification

Current stage as of 2026-06-12:

Physical Intelligence: S3/S4 research-platform and ecosystem evidence; not commercial deployment proof.
Skild AI: S4 capital + industrial-channel + data-flywheel evidence; not S5 scaled-commercial economics.
Overall model layer: early platform formation. Stronger than demo-only, weaker than revenue/margin-verified software layer.

7. Signal vs noise

Signal:

🟢 Public release of code/weights and reproducible tooling.
🟢 Cross-embodiment / cross-task evaluation with disclosed model/data approach.
🟢 Named industrial partners where deployment path is explicit.
🟢 Acquired installed robotics/orchestration assets that can generate real operational data.
🟢 Disclosed revenue, ARR, gross margin, customer retention, robot attach rate, or contract economics if it appears later.

Noise:

Viral robot videos without task distribution and failure data.
“Any robot, any task” language without deployment metrics.
Funding size alone.
Partner logos without contract economics.
Model benchmark demos that do not transfer into customer ROI.

8. What would upgrade this to S5

Upgrade criteria:

A model provider discloses paid deployment across multiple customers with robot counts, task hours, uptime, intervention rate, and contract value. 🟢 required evidence
A customer confirms ROI/payback or labor/productivity economics attributable to the model layer, not just hardware automation. 🟢 required evidence
Filings or audited disclosures show robotics software/model revenue becoming material with durable gross margin. 🟢 required evidence
An OEM discloses model licensing or revenue share with a third-party model provider at production scale. 🟢 required evidence
Repeat orders or expansion deployments show the model improves over time through data flywheel rather than requiring one-off engineering. 🟢 required evidence

9. Common misconceptions

Misconception 1: “Robot foundation model = LLM moment for robotics is already here.” Correction: The public evidence shows better generalization research, not yet a proven software revenue flywheel. Robotics requires physical reliability, safety, service cost, and customer ROI, not just model demos.

Misconception 2: “Open-source weights mean the model layer will be commoditized.” Correction: openpi may commoditize experimentation, but value may still sit in proprietary data, deployment tooling, safety validation, customer workflow integration, and fleet learning. This remains unresolved.

Misconception 3: “Big financing proves Skild is winning.” Correction: US$1.4bn Series C / >US$14bn valuation is a serious capital signal, but capital is not evidence of gross margin, retention, or scaled deployment economics.

Misconception 4: “Partnership with ABB/UR/NVIDIA means deployment economics are proven.” Correction: partnership is S4 channel evidence; S5 needs robot counts, contract economics, uptime/intervention data, and customer ROI/payback.

10. Think Deeper questions

In robotics, is the scarce asset the model architecture, the real-world data stream, or the deployment workflow integration?
Will OEMs prefer vertical model stacks for safety/reliability reasons, or will horizontal model providers win because data generalizes across embodiments?
What is the first observable financial metric for the model layer: software ARR, per-robot licensing, outcome-based pricing, or embedded OEM economics?
Does open-source tooling accelerate the whole category while reducing individual model-provider pricing power?

11. Public-site draft section

The missing layer: robot foundation models

Robotics is not only a hardware cost curve. The harder question is whether robots can generalize outside scripted demos.

Two public evidence curves are worth tracking:

Physical Intelligence: π0 → openpi → π0.5 → π0.7 shows the research curve from generalist robot policy to open tooling and broader generalization claims. Strong evidence of model-layer progress; weak evidence so far on commercial economics.
Skild AI: omni-bodied “one brain” strategy, US$1.4bn Series C at >US$14bn valuation, ABB/UR/NVIDIA partnerships, and Zebra/Fetch robotics asset acquisition show a capital + deployment-channel curve. Strong S4 signal; still missing S5 proof.

The important distinction: a better model can change the commercialization curve only if it reduces deployment cost, intervention rate, failure rate, and payback time in customer workflows. Until those metrics appear, model progress is a powerful signal — but not yet proof of a return cycle.

Footer caveat: evidence map only. No company ranking. No trade recommendation. Model claims require deployment economics before S5.

12. Source list

Physical Intelligence homepage, accessed 2026-06-12. https://www.pi.website/ 🟢
Physical Intelligence, “π0: Our First Generalist Policy,” published 2024-10-31. https://www.pi.website/blog/pi0 🟢
Physical Intelligence, “Open Sourcing π0,” published 2025-02-04. https://www.pi.website/blog/openpi 🟢
Physical Intelligence, “π0.5: a VLA with Open-World Generalization,” published 2025-04-22. https://www.pi.website/blog/pi05 🟢
Physical Intelligence, “π0.7: a Steerable Model with Emergent Capabilities,” published 2026-04-16. https://www.pi.website/blog/pi07 🟢
Skild AI homepage, accessed 2026-06-12. https://www.skild.ai/ 🟢
Skild AI, “Announcing our $300M Series A Funding,” published 2024-07-09. https://www.skild.ai/blogs/announcing-our-300m-series-a 🟢
Skild AI, “Building the general-purpose robotic brain,” published 2025-07-29. https://www.skild.ai/blogs/building-the-general-purpose-robotic-brain 🟢
Skild AI, “The case for an omni-bodied robot brain,” published 2025-09-24. https://www.skild.ai/blogs/omni-bodied 🟢
Skild AI, “Learning by watching human videos,” published 2026-01-12. https://www.skild.ai/blogs/learning-by-watching 🟢
Skild AI, “Announcing Series C,” published 2026-01-14. https://www.skild.ai/blogs/series-c 🟢
Skild AI, “The Reindustrial Revolution: Partnering with ABB Robotics, Universal Robots, and NVIDIA,” published 2026-03-19. https://www.skild.ai/blogs/reindustrial-revolution 🟢
Skild AI, “Skild AI Acquires Zebra Technologies' Robotics Arm to Bring Omni-Bodied Intelligence to Warehouses,” published 2026-04-15. https://www.skild.ai/blogs/skild-zebra 🟢

13. Public safety flag

Public-safe if kept as an evidence map. Do not include Hugo private portfolio data. Do not frame PI, Skild, NVIDIA, ABB, Universal Robots, Zebra, Tesla, Figure, Unitree, or any supplier as buy/sell/hold. Do not imply model-provider value capture without revenue/margin or contract evidence.