Physical Intelligence Market Research

1. Introduction

Physical Intelligence (abbreviated π / PI) differentiates itself within this group as a "pure-brain model vendor": it is the only company among the group's ten that does not build hardware at all, positioned as the "robot-native foundation model layer" — summarized by the industry as "ChatGPT for robots": one model driving any hardware. The other nine companies in the group, regardless of their routes, all earn revenue by "selling robots" or "selling robot solutions"; PI only sells (or opens) models, leaving the hardware to Unitree, Figure, and robotics companies worldwide.

1.1. Company Overview

1.2. Development History and Funding Lineage

1.3. Business Model

PI ships no complete machines and has no unit revenue: its business model is model licensing / partnerships + open-source ecosystem. "The Physical Intelligence Layer" (official blog, 2026-02-24) articulates its platform ambition — to serve as the intelligence layer for all robotics companies. The open-source portion (Apache license) functions as a de facto developer acquisition funnel, while the closed-source latest weights (π*0.6 / π0.7) preserve commercialization headroom. The Series C funding status must be presented per the "reported under negotiation" framing and must not be written as closed.

2. Glossary

3. Feature Description and Model Lineage

3.1. π Series Model Lineage

3.2. openpi Open-Source Strategy

openpi's (Apache license) open-source boundary is clear: π0 / π0-FAST (2025-02-04, including 7 fine-tuned checkpoints) and π0.5 weights (2025-09, including LIBERO / DROID variants and native PyTorch) are open; π*0.6 and π0.7 weights are not open-sourced. The repository has ~12,800 stars as of mid-2026. This "old-generation open, new-generation closed" strategy simultaneously achieves ecosystem positioning and commercial protection — see the governance analysis in Section 5.6.

4. Platform Architecture

4.1. Pure Model Layer Tech Stack

PI's product is the model stack itself: an end-to-end perception-language-action VLA (the model core of L1/L2), delivering 50Hz-level real-time control output via flow matching action chunks; MEM provides cross-time memory; RECAP and RL Tokens provide experience learning. Its architectural stance is: converge intelligence into one model and leave diversity to the hardware — forming, together with Figure (model-hardware bundling) and NVIDIA (GR00T platform + ecosystem), three "brain supply" models.

4.2. Data Strategy

π0's foundation is 10,000+ hours, 7 embodiments, 68-task cross-embodiment real-machine data; π0.5 introduces internet multimodal knowledge (knowledge-insulated co-training); π*0.6 / RL Tokens then shift to self-collected experience data (generated by the robot's own operation, consumed by online reinforcement learning), evolving data acquisition from "manual collection" toward "collection-by-operation". PI owns no hardware, so its data strategy must rely on partner networks and open-source community feedback — an inherent constraint of the pure-model route.

4.3. π Model Lineage Evolution Diagram

Figure 8-1 | π series model lineage evolution (2024-10 — 2026-04, per official blog)

Data source: verified item-by-item against Physical Intelligence's official blog (pi.website), information current as of 2026-09-12.

5. Harness Design

5.1. L1 Context Engineering Layer

The π series end-to-end assembles visual and language context into action policies; knowledge-insulated co-training (π0.5) resolves the conflict between the two kinds of context — "internet knowledge vs. robot action data" — a rare explicit context-engineering innovation in the embodied domain; steerability (π0.7) allows guiding the model's behavior at inference time, equivalent to a runtime context-control interface.

5.2. L2 Tools and Execution Layer

flow matching action chunks output continuous control at up to 50Hz, and Real-Time Action Chunking with Large Models (2025-06) specifically addresses maintaining accuracy and speed under high latency — a direct engineering investment in the L2 real-time control loop. Cross-embodiment data (7 embodiments) ensures the model's portability as a "general-purpose execution core". PI has no hardware of its own, so the other half of L2 (hardware adaptation) is done by customers.

5.3. L3 Orchestration and Control Layer

π0.5's 10–15 minute multi-stage household tasks and the long-horizon tasks over 10 minutes supported by MEM show that long-horizon task orchestration is borne by the model + memory rather than an external workflow engine. PI has not released agent-operating-system-type components (compare Limit AI's COSA); its stance can be summarized as: orchestration capability should live in the model.

5.4. L4 Memory and State Layer

MEM (Multi-scale Embodied Memory, 2026-03) makes PI the only vendor in this group to explicitly publish a memory mechanism with paper-grade blog posts: multi-scale long- and short-term memory lets the model maintain state consistency across tasks over 10 minutes. This is the hardest public evidence for the L4 layer.

5.5. Evaluation and Observation Layer

π0.5 has become a common VLA baseline in academia: papers such as TurboVLA, N0-VTLA, and LENS use it as a reference baseline — effectively being "passively benchmarked" by global academic evaluation. RECAP and RL Tokens, meanwhile, build the online evaluation-improvement loop into the training recipe (doubled throughput; mastering sub-millimeter tasks in 15 minutes per official claims). On honest disclosure, the academic framing noted a limitation of 0/10 task success when no verification pipeline exists, and PI did not shy away from failure cases. π0.7 is officially said to "rival expert fine-tuned models" with no third-party benchmark endorsement; judgment should be reserved.

5.6. L6 Governance and Security Layer

PI's L6 evidence is mainly open-source governance: the Apache license plus a clear open-source boundary (which weights are open, which are held back) is itself a form of capability-diffusion governance; open-sourcing older generations reduces single-point dependency risk, while closing newer generations protects commercial and potential misuse surfaces. Runtime safety governance for robots (emergency stop, behavior guardrails) is outside its responsibility domain — borne by hardware vendors and integrators — and this is a structural L6 gap of the pure-model route.

5.7. Six-Layer Strength Summary

6. Real-world Cases

6.1. π0.5's Academic Baseline Status

π0.5 (4B parameters, runnable on a single RTX 4080 Super) has become a common reference for VLA research in academia: subsequent papers such as TurboVLA, N0-VTLA, and LENS all use it as their baseline model. For a company founded only two years ago, "being benchmarked" is a more genuine indicator of industry standing than revenue — it means robotics researchers worldwide default to measuring progress against π's models.

6.2. Partner Deployment Directions

The official blog post "The Physical Intelligence Layer" (2026-02-24) shows the real-world problems partners are already solving (such as logistics), but it does not disclose named customers, deployment unit counts, or contract amounts, marked [To be filled]. Together with Skild AI (~$14B valuation) and Figure (hardware + model bundling), it forms a competitive comparison among three "robot brain" supply routes.

6.3. Capital Status and Industry Position

Funding lineage: Seed $70M (2024-03) → Series A $400M / $2.4B valuation (2024-11, led by Bezos) → Series B $600M / $5.6B valuation (2025-11-20, led by CapitalG / Google), cumulatively over $1.1B raised. Series C of $1B / $11B valuation has reportedly been under negotiation since 2026-03 (Founders Fund expected to lead, Lightspeed in talks), not closed as of the information cutoff date. NVIDIA's GTC 2026 listed Physical AI among the five strategic pillars and presented a trillion-dollar market narrative, providing macro tailwind for this track. It should be noted that circulating claims such as "Anthropic acquiring PI" have no first-hand source backing, are prohibited material, and are not adopted as facts in this document.

7. Summary

7.1. Strengths and Limitations

Strengths: the π lineage's methodological completeness leads globally (VLA architecture → tokenization → open world → RL → memory → steerability, six-level jump in two years); the academic baseline status attracts talent and ecosystem pull; the Apache open-source strategy secures de facto standard positioning; and the asset-light model that does not carry hardware.

Limitations: having no hardware makes data acquisition dependent on external sources, and the scale of cross-embodiment data is constrained by the partner network; there is no unit revenue, and the commercialization path (license pricing, SLA) is unvalidated; there is funding and valuation uncertainty during the unclosed Series C; and runtime safety is outside its domain, leaving end-to-end delivery responsibility ambiguous.

7.2. Applicability Boundaries

Three types of buyers fit: robotics companies that own hardware but lack a "brain" (model licensing / collaborative R&D); the academic and open-source community (openpi is directly usable); and research institutions that need a VLA baseline for evaluation. It does not fit end-factory customers wanting "turnkey complete machine solutions" — those need vendors like Figure, UBTech, or AgiBot.

7.3. Selection Recommendations

1. When hardware vendors select PI, they should focus on evaluating hardware adaptation costs: the data requirements for cross-embodiment fine-tuning, and the matching of the 50Hz control loop with the hardware's real-time bus.
2. Using openpi (π0.5) for PoC and negotiating commercial licensing for closed-source new models is a reasonable two-stage path today.
3. Monitor the Series C closing status and the first named commercial customers — until these two things materialize, PI's commercialization narrative should be treated as "pre-commercialization".
4. When citing π series success-rate data, clarify whether it is an academic-benchmark framing or a customer-scenario framing; the two must not be conflated.

Information Gap Statement

1. Series C status: $1B / $11B valuation is "reported under negotiation", not closed as of 2026-09-12; this document strictly presents it per the negotiation framing.
2. Named commercial customers and license pricing are not disclosed, marked [To be filled].
3. Original paper (arXiv) links for π0.5 / π*0.6 / π0.7 were not individually obtained in this search; this document cites pi.website's official blog as authoritative (release dates verified item-by-item).
4. π0.7 "rivals expert fine-tuned models" is an official claim with no third-party benchmark endorsement.
5. "Anthropic acquiring PI" is a D-grade rumor (no first-hand source), handled as prohibited material and not written into the body.

8. References

1. π0: Our First Generalist Policy — Physical Intelligence official blog, 2024-10-31. https://www.pi.website/blog/pi0
2. Open Sourcing π0 — Physical Intelligence official blog, 2025-02-04. https://www.pi.website/blog/openpi
3. π0.5: a VLA with Open-World Generalization — Physical Intelligence official blog, 2025-04-22. https://www.pi.website/blog/pi05
4. π*0.6: a VLA that Learns from Experience — Physical Intelligence official blog, 2025-11-17. https://www.pi.website/blog/pistar06
5. π0.7: a Steerable Model with Emergent Capabilities — Physical Intelligence official blog, 2026-04-16. https://www.pi.website/blog/pi07
6. VLAs with Long and Short-Term Memory (MEM) — Physical Intelligence official blog, 2026-03-03. https://www.pi.website/research/memory
7. Precise Manipulation with Efficient Online RL (RL Tokens) — Physical Intelligence official blog, 2026-03-19. https://www.pi.website/research/rlt
8. The Physical Intelligence Layer (partner directions) — Physical Intelligence official blog, 2026-02-24. https://www.pi.website/blog/partner
9. OpenPI — AI Wiki (openpi timeline / open-source boundary / funding), 2026. https://aiwiki.ai/wiki/openpi
10. Physical Intelligence's $11B Bet — AgentMarketCap, 2026-04-17. https://agentmarketcap.ai/blog/2026/04/17/physical-intelligence-11b-robot-foundation-model-vla-thesis
11. Physical Intelligence — Silicon Valley Invest Club (funding table), 2026-08. https://siliconvalleyinvestclub.com/physical-intelligence/
12. Physical Intelligence Eyes $11B Valuation — AI2.work, 2026. https://ai2.work/blog/physical-intelligence-eyes-11b-valuation-in-ai-robotics-mega-round
13. Humanoid Robot Imitation Learning Companies (π0.5 baseline status) — Teahose, 2026-08. https://www.teahose.com/themes/humanoid-imitation-learning