Robotics foundation models and embodied AI

Market Map: Robotics Foundation Models & Embodied AI

(1) Thesis

The robotics stack is being restructured from vertically integrated, task-specific automation into a horizontal AI layer, where value migrates from hardware to whoever owns the generalist "robot brain" and the data flywheel that trains it—mirroring the LLM-to-application dynamic in software.

---

(2) Segments

A. Foundation Model / "Robot Brain" Labs

Building generalist vision-language-action (VLA) models intended to run across many embodiments.

• Physical Intelligence (Pi/π) — π0 model; the archetypal hardware-agnostic play
• Skild AI — general-purpose robot brain, CMU-rooted
• Google DeepMind — RT-2, RT-X, Gemini Robotics
• NVIDIA — GR00T foundation models + Isaac ecosystem
• Covariant — early mover (RFM-1), team largely absorbed by Amazon in 2024

Dynamics: Massive capital raises pre-revenue; the core bet is that cross-embodiment generalization works. Data scarcity (no "internet of robot actions") is the binding constraint, driving frenetic competition for teleoperation and simulation data.

B. Humanoid Full-Stack Companies

Vertically integrated hardware + model, betting the humanoid form factor unlocks human environments.

• Figure AI — Helix model; broke off OpenAI partnership to go in-house
• Tesla (Optimus) — leverages FSD data infrastructure
• 1X Technologies — NEO, home-focused; OpenAI-backed
• Agility Robotics — Digit; furthest along in actual warehouse deployments
• Unitree — Chinese cost leader, driving hardware prices down dramatically
• Apptronik — Apollo; Google DeepMind partnership

Dynamics: Capital-intensive arms race; pilots (BMW, GXO, Mercedes) are real but ROI unproven. Chinese entrants are commoditizing the hardware layer faster than Western players expected.

C. Simulation, Data & Training Infrastructure

The "picks and shovels" — synthetic data, sim-to-real, teleoperation pipelines.

• NVIDIA (Isaac Sim / Omniverse / Cosmos) — dominant position
• Scale AI — robotics data labeling/collection expansion (less certain of current scale of this line)
• Hillbot (less certain — early-stage sim-focused startup)
• Rerun — visualization/data tooling for robotics (smaller, dev-tools scale)

Dynamics: Thinner competitive field than expected; NVIDIA's gravitational pull is enormous. Real-world data collection remains artisanal — an underbuilt layer.

D. Applied Embodied AI (Vertical Deployment)

Applying learned policies to specific commercial workflows today.

• Amazon Robotics — internal deployment at massive scale; absorbed Covariant talent
• Dexterity — warehouse manipulation
• Ambi Robotics — parcel sorting with foundation-model approach (PRIME-1)
• Path Robotics — welding
• Collaborative Robotics (Cobot) — non-humanoid mobile manipulation
• Intrinsic (Alphabet) — industrial robotics software platform

Dynamics: The only segment with meaningful revenue today. Trades ceiling for near-term unit economics; risk is being leapfrogged by generalist models.

E. Autonomy Incumbents Repositioning as Embodied AI

• Waymo, Wayve, Boston Dynamics (Hyundai) — Wayve explicitly frames its models as embodied AI foundations; Boston Dynamics partnering with Toyota Research Institute on LBMs.

Dynamics: Deep pockets, proven real-world data pipelines; question is whether driving/legged data transfers to manipulation.

---

(3) Where Value Accrues

Segment A (foundation model labs) wins the most value — if generalization works — but Segment B captures it if it doesn't. The structural argument for A: hardware is commoditizing (Unitree's pricing proves this), so durable margin lives in the model layer, licensed across embodiments like an OS. The counterargument, and why Figure/Tesla went in-house: unlike LLMs, robotics data is generated by the hardware, so the vertically integrated player controls the data flywheel. My call: value bifurcates — one or two full-stack humanoid winners (B) plus one dominant model licensor (A), with NVIDIA (C) as the near-certain toll-taker regardless of outcome. If forced to pick one segment: A, because the model layer's TAM spans all form factors while any single hardware bet does not.

(4) White Space

1. Real-world robot data marketplace / "data union" for deployed fleets. Millions of existing industrial robots (Fanuc, ABB, KUKA installs) generate trajectory data that is discarded or siloed. No one operates a neutral aggregation, licensing, and rev-share layer that turns brownfield fleets into training data suppliers for foundation model labs — the Scale AI of action data, sourced from production rather than teleoperation farms.
2. Safety certification, evals, and insurance infrastructure for learned policies. Regulators and insurers have no framework for stochastic neural policies controlling physical force near humans. A company doing standardized red-teaming, runtime guardrails, and actuarially-backed certification for embodied AI (an "UL Labs + eval harness" for robots) would sit at a mandatory chokepoint for every deployment in Segment B and D. Nothing comparable exists today beyond nascent academic benchmarks.