Which companies build foundation models for robotics, and how do they compare?

The brain-provider cohort

The foundation-model-for-robotics category operates in 2026 across multiple distinct strategic theses. Each major brain-provider company pursues a structurally different bet on how foundation models for robotics should be built, deployed, and commercialized:

• Skild AI: cross-platform general-purpose brain thesis
• Physical Intelligence: research-publication VLA models (Pi-0, Pi-0.5)
• Covariant: warehouse-automation foundation specialization
• Google DeepMind: Gemini Robotics + RT-2 across AV and humanoid research
• OpenAI Robotics: relaunched May 2026 after the 2021 Dactyl team dissolution
• NVIDIA Project GR00T: cross-platform humanoid integration aligned with NVIDIA stack
• Meta AI (FAIR + Reality Labs): research-publication emphasis

The cohort operates at the research-and-demonstration verification tier of DEPLOY's four-tier capability framework. Commercial deployment at the depth that humanoid OEMs (Figure 03 at BMW Spartanburg; Agility Digit at GXO Flowery Branch) have achieved has not landed at comparable depth in the brain-provider tier.

Per-company strategic comparison

Skild AI

Pittsburgh-based with CMU robotics heritage. The cross-platform general-purpose brain thesis aims at a single VLA model deployable across multiple humanoid + quadruped + manipulator platforms. Strategic position: brain-provider value increases if cross-platform thesis pays out; humanoid OEM brain-development internal effort decreases correspondingly.

See the full Skild AI explainer for institutional facts, funding context, and detailed strategic positioning.

Physical Intelligence

UC Berkeley research lineage through Sergey Levine's robotics research lab. The Pi-0 (introduced 2024) and Pi-0.5 (subsequent iteration) VLA models emphasize transformer-based architecture with multi-modal training data. Research-publication depth distinguishes Physical Intelligence from peer brain providers that emphasize commercial-deployment depth more aggressively.

The verification surface: published papers documenting capability claims; benchmark scores on academic task suites; partnership integration depth at varying disclosure. Per DEPLOY's verified-vs-claimed framework, the research-publication depth is verifiable; commercial-scale deployment at the depth humanoid OEMs have achieved is the forward verification surface.

Covariant

UC Berkeley research lineage through Pieter Abbeel. Covariant's specialization in warehouse automation distinguishes it from cross-platform peers (Skild) and research-publication peers (Physical Intelligence). The warehouse-automation focus produces specific commercial relationships (AWS partnership context per disclosure) and specific verification surfaces (warehouse-task task-completion rates rather than cross-platform general-purpose evaluation).

Covariant's competitive position is interesting because warehouse-automation specialization may produce earlier commercial-scale verification than general-purpose cross-platform claims. The verification trajectory for warehouse-specific brains may diverge from general-purpose brains over the next several years.

Google DeepMind

Large research organization with multiple robotics research lines including RT-2 (the Robotics Transformers paper line that produced foundational VLA architecture insights) and Gemini Robotics (extending Gemini multimodal models to embodied action). DeepMind's institutional position differs from pure-play brain providers: research depth + integration with Google product surfaces (Waymo for AV; Google Cloud for compute + deployment infrastructure) + brand recognition from broader Gemini work.

Commercial integration depth in robotics specifically is shaped by Google's broader strategic position rather than dedicated robotics commercial-deployment focus. For the broader frontier-AI-lab cluster context including DeepMind alongside OpenAI and Meta, see frontier AI labs entering robotics.

OpenAI Robotics

Relaunched May 31, 2026 after the Dactyl team dissolved around 2021. Current scope is developing under Aditya Ramesh's leadership (Worldsim research line continuation toward embodiment). OpenAI's existing minority investments in 1X Technologies (2023) and Figure AI (2024) bridged the 2021-2026 hiatus; the May 2026 internal program represents the first formal OpenAI robotics organization since Dactyl.

Verification posture per registry tie-in indexability gate: OpenAI Robotics entity is accessible at direct URL but pending verifiable-robot-platform reference to surface in registry paginated listing. The cap-flag is the editorial truth for an emerging-but-yet-to-ship internal robotics program.

NVIDIA Project GR00T

Cross-platform humanoid foundation model aligned with NVIDIA's broader robotics stack (Isaac Sim simulation; Jetson edge hardware; Omniverse 3D infrastructure). NVIDIA's strategic position is distinctive: the company sells the underlying compute and simulation infrastructure that brain-provider tier work depends on, plus operates its own brain-provider tier product. The vertical integration shapes commercial relationships across the brain-provider tier.

Meta AI (FAIR + Reality Labs)

Research-publication emphasis via Facebook AI Research (FAIR) and Reality Labs (consumer AR/VR product line). Meta's institutional position favors research depth and consumer-product integration (Reality Labs glasses + future consumer-AI products) over commercial robotics deployment. The brain-provider tier work at Meta operates at research-and-publication scale; commercial humanoid integration is not the primary strategic focus.

What the framework verifies and what it does not

Per DEPLOY's verified-vs-claimed framework, the brain-provider tier cohort's current verification posture is structurally distinct from humanoid OEM tier verification:

• Research output verified across cohort: published papers + benchmark scores + demonstration videos document VLA capability at research-and-demonstration scale.
• Cross-platform deployment claimed at varying depth: pure-play brain-provider strategies (Skild, NVIDIA GR00T) emphasize cross-platform; research-focused approaches (Physical Intelligence, Meta) emphasize publication depth; specialized approaches (Covariant) focus on specific verticals.
• Commercial-scale deployment lags humanoid OEM tier substantially: BMW Spartanburg / GXO Flowery Branch / Mercedes-Benz pilots / Catalyst Brands Reno deployments anchor humanoid OEM verification. Brain-provider tier integration-partner deployments at comparable customer-facility verification depth have not landed at the same disclosure cadence.
• Cap-flag: per-model commercial-deployment counts; per-partner integration depth; cross-platform transfer at commercial scale; financial sustainability of brain-provider tier business models. All currently sit at cap-flag tier; the cap-flag is the editorial truth, not a gap.

Where to go for context

For the structural framework distinguishing brain-provider tier from OEM-platform tier, see brain-provider tier vs OEM-platform tier distinction. For the foundation-model-for-robotics category background, see what is a foundation model for robotics.

For broader frontier-AI-lab cluster context including OpenAI + DeepMind + Meta robotics work, see frontier AI labs entering robotics. For the OpenAI Robotics relaunch specifically, see the foundational signal.

For canonical institutional depth at the registry layer, see the per-company records: Skild AI, Physical Intelligence, Google DeepMind, OpenAI Robotics. For methodology canonical references applicable to brain-provider landscape comparison: captive vs third-party brain providers (CANONICAL cross-provider integration model gradient) + the 4-way autonomy-boundary taxonomy (brain-provider integration intersects autonomy-boundary at deployment depth) + the 9-tier source-quality rubric.