ExplainersBrain providers & foundation models
Brain providers & foundation models
Robotics infrastructure tier. Companies building the AI that humanoid OEMs, AV operators, and drone makers integrate as embodied control.
19 explainers
Brain providers are the robotics infrastructure tier. The cohort builds the foundation models, autonomy stacks, and embodied control software that humanoid OEMs, AV operators, drone makers, and adjacent physical AI cohorts integrate as their core AI substrate rather than building from scratch. The platform-builder vs brain-provider distinction is editorially substantive: humanoid OEMs (Figure, Apptronik, Agility, 1X, Boston Dynamics, Tesla) and AV operators (Waymo, Tesla, Aurora) build hardware + integrate AI + sell as commercial product; brain providers (Skild AI, Physical Intelligence, NVIDIA GR00T, Wayve, Covariant RFM-1, Dyna Robotics DYNA-1, Google DeepMind Gemini Robotics, OpenAI Robotics, frontier-lab equivalents) build AI + license or partner for integration. They do not build platform hardware.
The cluster's editorial throughline is the verification surface differential. Platform-builder commercial verification operates against deployed-platform evidence (Figure 02 at BMW Spartanburg's 30,000 vehicles; Agility Digit at GXO Flowery Branch's 100,000-tote scaled-throughput). Brain-provider commercial verification operates against integration-partner evidence (Skild integration depth across humanoid customers; Gemini Robotics integration across AV + humanoid partners) rather than deployed-platform evidence. Verification asks structurally different questions.
Foundation-model tier specialization carries the brain-provider cohort's structural distinctions per Agent A source-depth hygiene pass. Physical Intelligence anchors foundation-model-for-robotics ($5.6B confirmed valuation; pi0 + pi05_base open under Apache-2.0 + π0.6 + π0.7 closed; lab + limited-customer-pilot deployment). NVIDIA GR00T anchors foundation-model-for-physical-AI-general (N1 noncommercial NVIDIA OneWay → N1.7 commercial-open license evolution; NO verified production deployment cap-flag; BD + Unitree + Sharpa research-tier partners). Wayve anchors foundation-model-for-driving ($8.6B Series D Feb 2026 largest UK AI; AI Driver production + GAIA-3 world-model lineage; commercial via Stellantis STLA AutoDrive). Covariant RFM-1 anchors foundation-model-for-robotic-manipulation (corporate-state vs model-state separation; founders Amazon-acqui-hired August 2024; legacy Brain via KNAPP verified-deployment NOT RFM-1). Dyna Robotics DYNA-1 anchors foundation-model-for-dexterous-manipulation (pilot-not-scale service-industry deployment; founder lineage Gao + Yang + Ma per primary-source verification).
The frontier AI lab cluster (OpenAI's May 2026 robotics relaunch + Google DeepMind's Gemini Robotics + Meta AI's FAIR + Reality Labs work) sits at the brain-provider tier. The institutional commitments are real; the integration-partner deployment at commercial scale is the forward question. Per the no-pre-load-from-single-instance discipline, three cases (OpenAI + DeepMind + Meta) cleared the threshold for codifying the cluster as editorially substantive.
Brain providers operate at registry-only depth: brains are not consumer-purchasable products by design, so there is no consumer surface for foundation models. The canonical enumeration lives at the registry's /brains category where the always-current brain list surfaces. Brain entities downstream cluster integration: humanoid VLA models (helix for Figure; gr00t-n1; pi0 + pi05 for Physical Intelligence; skild-brain for Skild AI; gemini-robotics for Google DeepMind); AV driver stacks (waymo-driver + zoox-driver + aurora-driver + tesla-fsd-bot); defense autonomy stacks (lattice for Anduril; hivemind for Shield AI; skydio-autonomy-engine for commercial drones). Standalone brain-provider company entities: Skild AI + Physical Intelligence + NVIDIA + Wayve + Covariant + Dyna Robotics.
For the framework canonical reference + canonical worked examples demonstrating the discipline operationally, see how DEPLOY verifies. For the canonical category umbrella that includes brain providers alongside the other physical AI cohorts, see what is physical AI.
For methodology pillar canonical references applicable to the brain-provider cohort: captive vs third-party brain providers (CANONICAL: captive-vs-third-party integration gradient as foundational brain-provider framework); the 4-way autonomy-boundary taxonomy (brain-provider integration intersects autonomy-boundary classification at deployment depth); the 9-tier source-quality rubric (brain-provider partnership claims source classification).
Adjacent clusters
- Humanoid robots: Humanoid platforms integrate foundation models (Figure Helix VLA + Apptronik partnerships + 1X learned-control stack); the humanoid cluster surfaces the platform-builder side.
- Robotaxis & autonomous vehicles: AV foundation models (DeepMind Gemini Robotics for AVs; in-house FSD stacks for Tesla) operate at brain-provider tier integrated into AV platforms.
- Autonomous drones: Anduril Lattice + Shield AI Hivemind operate as defense-context autonomy stacks adjacent to humanoid VLA foundation models; structurally parallel brain-provider position.
Featured
What is Physical Intelligence?
Physical Intelligence (PI) is a foundation-model-for-robotics company building cross-embodiment AI brains. $5.6B valuation confirmed per Agent A source-depth hygiene (NOT $11B from aggregator drift; the $10B/$38B figures belong to Project Prometheus separate Bezos lab). Open-weights model lineage: pi0 (Apache-2.0) + pi05_base (verified open under Apache-2.0 per openpi README). Closed-weights: π0.6 + π0.7. Series B $400M March 2024 (current state worth re-verifying per Agent A note on ~$600M Series B context). Per DEPLOY's brain-providers cluster framework, PI anchors the foundation-model-for-robotics tier with lab-deployment in limited customer pilots.
Read →
What is NVIDIA GR00T?
NVIDIA GR00T is NVIDIA's foundation-model-for-physical-AI-general brain provider. Model lineage: N1 (original weights noncommercial under NVIDIA OneWay license) → N1.7 (commercial-open). KEY FRAMING per Agent A source-depth hygiene: NO verified production GR00T deployment exists anywhere (substantial cap-flag against aggregator narratives of 'GR00T at scale'). Partner matrix verified: only Boston Dynamics + Unitree + Sharpa partner-confirmed (all research-tier); Apptronik aggregator-inflated (primary brain is Gemini); Figure is competitor not deployment. Per DEPLOY's brain-providers cluster framework, GR00T anchors the foundation-model-for-physical-AI-general tier with research-tier deployment maturity per verified evidence.
Read →
What is Wayve?
Wayve is a UK-based brain-provider building foundation-model-for-driving for autonomous vehicles. Per Agent A source-depth hygiene: $8.6B valuation following Series D February 2026 (largest UK AI Series D on record). Model lineage: AI Driver (production end-to-end driving model deployed via OEM integration) + GAIA-3 (generative world model for driving simulation). Production deployment via Stellantis STLA AutoDrive partnership: commercial-tier deployment maturity (rare in the foundation-model brain-provider cohort). Cohort positioning: foundation-model-for-driving tier distinct from foundation-model-for-robotics (PI) + foundation-model-for-physical-AI-general (NVIDIA GR00T) + foundation-model-for-robotic-manipulation (Covariant) cohort tiers.
Read →
What is Dyna Robotics DYNA-1?
Dyna Robotics DYNA-1 is the foundation-model-for-dexterous-manipulation from Dyna Robotics, founded by Lindon Gao + York Yang + Jason Ma per Agent A source-depth hygiene pass. CRITICAL CAP-FLAG: pilot-not-scale deployment posture; no named production customers per verified evidence. Pilot deployments operate in service industries (hotels + restaurants + laundromats + gyms) NOT factories. Model availability: closed weights + closed code + no published paper per Agent A discipline. Cohort positioning: foundation-model-for-dexterous-manipulation tier distinct from foundation-model-for-robotics (PI) + foundation-model-for-physical-AI-general (NVIDIA GR00T) + foundation-model-for-driving (Wayve) + foundation-model-for-robotic-manipulation (Covariant) cohort tier positions.
Read →
All explainers in brain providers & foundation models
How did DEPLOY catch the 1X Redwood brain-vs-hardware framing error?
A dispatch framed 1X Redwood as '1X's industrial sibling to NEO consumer' with 'industrial-pilot with verified pilot deployments' (a humanoid hardware product). Audit-first verification against the registry surface showed Redwood is registered at /brains/1x-redwood as a brain entity (foundation model + captive AI substrate), NOT a humanoid hardware product. The 1X Technologies actual humanoid product lineup at the registry surface: /models/1x-eve (industrial) + /models/1x-neo (consumer). No /models/1x-redwood exists. The correction was caught before the entity anchor was authored, surfaced to the operator, and corrected at the editorial framing layer. The brain-entity vs hardware-model entity-type discipline is now editorially canonical. This piece documents the catch as a framework-in-action worked example demonstrating audit-first discipline + dispatch-framing-vs-codebase-reality verification + cap-flag honesty operating at editorial-anchor depth.
How did DEPLOY correct the Covariant corporate-state narrative?
Aggregator coverage of [Covariant](/explainers/what-is-covariant-rfm-1) frequently asserts the company was acquired by Intrinsic (Alphabet's industrial robotics arm). Per Agent A primary-source verification, the framing is wrong. Covariant operates continuing-but-diminished post-Amazon-AGI departure August 2024. Founders Pieter Abbeel + Peter Chen + Rocky Duan + ~25% of staff moved to Amazon's AGI team; Covariant the legal entity continues operating with Ted Stinson as CEO; legacy Covariant Brain continues verified commercial deployment via the KNAPP warehouse-automation partnership; non-exclusive license and IP retained by Covariant. The within-corporate-state framing is editorially distinctive: Covariant is simultaneously verified-continuing (corporate entity + CEO + commercial deployment via KNAPP) and verified-diminished (founder + leadership + capacity moved to Amazon AGI; reduced staff; research-tier RFM-1 model future cap-flagged). Two true claims at once; aggregator framings collapse them into a single wrong claim ('Covariant acquired by Intrinsic' or 'Covariant defunct after Amazon acqui-hire'). This piece documents the catch as framework-in-action worked example: corporate-state-precision discipline + within-entity verified-continuing-and-verified-diminished framing + acquirer-identification verification operating at editorial-anchor depth.
How did DEPLOY correct the Physical Intelligence valuation conflation?
Aggregator coverage of [Physical Intelligence](/explainers/what-is-physical-intelligence) frequently cites the company's valuation at $11B. Per Agent A primary-source verification: Physical Intelligence is valued at $5.6B confirmed. The $10B and $38B figures often cited in aggregator coverage belong to Project Prometheus, a separate Jeff Bezos-backed lab commonly conflated with Physical Intelligence due to overlapping investor coverage + adjacent positioning. The conflation operates as a worked example of entity-distinction discipline: two distinct entities (Physical Intelligence + Project Prometheus) with structurally similar funding patterns (Bezos backing) and adjacent positioning (robotics-adjacent AI research) get collapsed into one inflated valuation claim. Per [DEPLOY's framework discipline](/explainers/how-deploy-verifies), audit-first verification against named-entity primary sources + cap-flag transparency against the conflation produced the corrected $5.6B attribution. This piece documents the catch as framework-in-action worked example: entity-distinction discipline + primary-source-anchored funding verification + adjacent-entity-conflation rejection operating at editorial-anchor depth.
How does DEPLOY apply verified-vs-claimed at within-entity granularity?
The verified-vs-claimed framework operates at feature-granularity within a single device, not just device-granularity across the cohort. Two canonical worked examples surface the discipline operationally. THE WHOOP BPI SAGA: the Blood Pressure Insights feature on Whoop devices carries manufacturer-position-vs-FDA-position dispute; the July 14, 2025 FDA Warning Letter; a class action filed late 2025/early 2026; the dispute remains UNRESOLVED as of May 2026. The Whoop device exists at verified-product-feature depth; the BPI feature operates at claimed-health-monitoring-capability depth with unresolved FDA dispute. Same device, two structurally distinct verification states at the feature layer. THE APPLE WATCH SPO2 SAGA: the SpO2 (blood oxygen) feature on Apple Watch was removed from US-distributed units in January 2024 following ITC injunction in the Masimo patent dispute; restored in some configurations via subsequent design-around; verification posture varies by date + jurisdiction within the same device. Apple Watch hardware verifies consistently; SpO2 feature operates at structurally distinct verification states across US-pre-injunction + US-post-injunction + US-restored + ex-US jurisdictions. Architectural significance: the framework operates at feature-granularity, not just device-granularity. Within-entity verification states emerge when feature-specific regulatory disputes + patent litigation + jurisdiction-specific availability operate within the same hardware platform.
How does DEPLOY classify autonomy boundaries across physical AI?
Autonomy boundaries across physical AI sort structurally into a four-way taxonomy: autonomous-execution (machine executes the operation under AI control with operator supervision); AI-augmented operator-controlled (operator executes; software plans + tracks + bounds); replacement-robotics teleoperated (operator teleoperates remote actuators via console); and assistive co-pilot (operator directly manipulates with augmentation). The taxonomy applied editorially across surgical (Monogram mBos autonomous-execution + Mako/CORI/ROSA AI-augmented + da Vinci/Hugo/Ottava/Versius teleoperated + Maestro assistive); drone (new-defense AI-first autonomous-mission profiles + legacy-prime AI-augmented + remotely-piloted teleoperated baseline); maritime (Ghost Shark autonomous-mission ASW + Saronic Corsair operator-supervised + REMUS family teleoperated/programmed-mission); humanoid (1X NEO operator-supervised + Tesla Optimus teleoperated demonstrations + Apptronik Apollo industrial-AI-augmented); sidewalk-delivery (Starship + Serve operator-supervised L4 + remote-teleoperator edge cases). Verification criteria + aggregator-drift patterns per tier surfaced editorially. Per [how-deploy-verifies methodology editorial](/explainers/how-deploy-verifies), the autonomy-boundary classification operates as canonical worked example of framework discipline operating recursively across cohorts at form-factor-cell granularity.
How does DEPLOY verify physical AI claims?
DEPLOY exists to distinguish verified facts from claimed facts in physical AI. The framework operates a four-tier verification protocol applied uniformly across registry, news, consumer, and methodology surfaces: every claim resolves to verified (primary-source-confirmed), stated (operator-asserted at source-quality posture), claimed (asserted without primary-source confirmation), or absence (honest-not-known). Cap-flag convention surfaces honest confidence limits transparently rather than hiding them. Source-quality tier rubric weights primary sources (SEC + FDA + IR + peer-reviewed) heaviest, secondary established sources substantially, secondary industry sources moderately, and aggregator content as context only. Canonical worked examples (RingConn vs Happy Ring + Stelo vs Lingo + Figure 03 BMW correction + Stryker Portage MI correction + Physical Intelligence valuation correction + Covariant corporate-state precision + Monogram autonomy-boundary distinction) demonstrate the discipline operationally across cohorts. The framework operates at form-factor-cell granularity (biometric ring sub-cohort + glucose-cell + surgical orthopedic sub-cohort + humanoid consumer-vs-industrial) and applies maturity-stage discipline (research/pilot/commercial/production) symmetrically across corporate-scale and emerging entrants. This is the verification protocol that makes DEPLOY a reference institutional partners can audit.
How does DEPLOY's 9-tier source-quality rubric classify verification evidence?
DEPLOY's source-quality rubric operates a 9-tier structural classification of verification evidence: 5 primary-source tiers (SEC filing + FDA database + government record + company IR + academic peer-reviewed); 2 secondary-source tiers (established publication + industry publication); 1 aggregator tier (Wikipedia + SEO sites + blog aggregators); 1 unverified tier. The rubric refines the 4-tier source-quality framework from [DEPLOY's how-deploy-verifies methodology editorial](/explainers/how-deploy-verifies) by distinguishing primary-source subtypes that operate at structurally distinct verification depths within the primary tier (SEC + FDA + government record + company IR + academic each verify against different document categories with different jurisdiction + scope + audit-trail properties). The architectural significance: source-quality operates per-claim, not per-source. The same source can carry verified factual claims (FOIA-anchored contract figures) and unverified allegation framings (management critiques + deployment-posture characterizations) simultaneously. The 9-tier rubric provides the framework substrate for per-claim verification at the source-quality depth where each claim actually resolves.
What is 1X Redwood?
1X Redwood is the captive-brain / foundation model from [1X Technologies](https://registry.deploy.report/companies/1x-technologies), the AI substrate that powers the [1X NEO](/explainers/is-1x-neo-autonomous-or-controlled-by-humans) consumer humanoid and 1X EVE industrial humanoid. CRITICAL ENTITY SCOPE per audit-first verification: Redwood is 1X's BRAIN entity (registered at /brains/1x-redwood), NOT a humanoid hardware product. Common framing confusion: trade-press coverage occasionally references 'Redwood' as if it were a separate 1X humanoid robot; the verified entity scope is the captive-brain foundation model that integrates into 1X's hardware platforms (NEO + EVE). Per [DEPLOY's brain-providers cluster framework](/explainers/brain-providers), Redwood anchors the captive-brain archetype within the brain-providers cluster, structurally distinct from third-party brain providers (NVIDIA GR00T + Skild AI + Physical Intelligence + Wayve + Covariant + Dyna). Per [consumer-vs-industrial humanoid sub-cohort architecture](/explainers/consumer-vs-industrial-humanoid-archetypes): consumer humanoid archetype operates more likely captive-brain integration (Redwood as canonical worked example) vs industrial humanoid archetype's more likely third-party-brain integration claims (cap-flagged per primary-source verification).
What is a foundation model for robotics?
A foundation model for robotics extends the large-language-model paradigm to physical-action prediction. Trained on robot demonstrations rather than internet text alone, these models output action sequences (motor commands, manipulator trajectories) rather than text tokens. The category is dominated by vision-language-action (VLA) architectures that take camera images plus optional language instructions as input and produce action tokens as output. Companies building these models constitute the brain-provider tier of the robotics value chain, distinct from the humanoid OEM tier that builds the physical platforms the models run on.
What is Covariant RFM-1?
Covariant RFM-1 (Robotics Foundation Model 1) is Covariant's foundation-model-for-robotic-manipulation, announced March 2024. CRITICAL DISTINCTION per Agent A source-depth hygiene: corporate-state vs model-state separation. Corporate-state: Covariant founders + ~25% staff Amazon-acqui-hired August 2024 (Pieter Abbeel + Peter Chen + Rocky Duan + others moved to Amazon AGI); Covariant continues as IP/legal entity. Model-state: RFM-1 has NO verified production deployment; the verified-deployment Covariant model is the legacy Brain (deployed via KNAPP warehouse-automation partnership), NOT RFM-1. Cohort positioning: foundation-model-for-robotic-manipulation tier distinct from foundation-model-for-robotics (PI) + foundation-model-for-driving (Wayve) + foundation-model-for-physical-AI-general (NVIDIA GR00T) + foundation-model-for-dexterous-manipulation (Dyna) cohort tier positions.
What is Skild AI and the Skild Brain foundation model?
Skild AI is a US foundation-model-for-robotics company headquartered in Pittsburgh, Pennsylvania, founded by Deepak Pathak and Abhinav Gupta (both Carnegie Mellon University robotics-research alumni). The company's Skild Brain product is a vision-language-action (VLA) foundation model architected around a cross-platform general-purpose thesis: a single model trained to operate across multiple robot platforms rather than platform-specific brains. Skild has raised substantial 2024-2025 funding rounds; the company is privately held and represents a distinctive position in the brain-provider tier of the robotics value chain.
What's the difference between captive and third-party brain providers in physical AI?
The brain-provider tier in physical AI splits structurally across two integration models with different counterparty-risk + cross-device-interoperability + competitive-dynamics implications. CAPTIVE BRAIN: vertically-integrated within a single corporate-state entity that also builds the hardware embodiment. The AI substrate and the platform ship together under the maker's full-stack control. [1X Redwood](/explainers/what-is-1x-redwood) is the canonical worked example. Redwood powers 1X NEO consumer + 1X EVE industrial under single corporate-state ownership. Other captive-brain candidates (Figure Helix on Figure 03; Tesla's FSD-derived stack on Optimus; Sanctuary AI cognitive architecture on Phoenix) operate at captive positioning but with verification-depth cap-flagged pending primary-source confirmation. THIRD-PARTY FOUNDATION-MODEL: licensed or partner-integrated AI substrate that ships across multiple hardware platforms. [Physical Intelligence](/explainers/what-is-physical-intelligence) pi0/pi05 (Apache-2.0 open) + π0.6/π0.7 (closed); [NVIDIA GR00T](/explainers/what-is-nvidia-groot) N1 noncommercial → N1.7 commercial-open; [Wayve](/explainers/what-is-wayve) AI Driver + GAIA-3 commercial via Stellantis; [Covariant RFM-1](/explainers/what-is-covariant-rfm-1) legacy Brain via KNAPP; [Skild AI](/explainers/what-is-skild-ai) integration-partner forward; [Dyna Robotics DYNA-1](/explainers/what-is-dyna-1) foundation-model-for-dexterous-manipulation. The captive-vs-third-party gradient is editorial signal at the cohort-architecture layer per [consumer-vs-industrial humanoid sub-cohort architecture](/explainers/consumer-vs-industrial-humanoid-archetypes): consumer humanoid archetype operates more likely captive-brain (Redwood canonical); industrial humanoid archetype more likely third-party-brain integration claims (cap-flagged per primary-source verification).
What's the difference between robotics brain providers and robot makers?
Robotics value chain operates across three structural tiers. Brain-provider tier companies (Skild AI, Physical Intelligence, Covariant, Google DeepMind, OpenAI Robotics, NVIDIA Project GR00T) build foundation models for robotics without making hardware. OEM-platform tier companies (Figure AI, Apptronik, 1X Technologies, Tesla, Agility Robotics, Boston Dynamics, Unitree, UBTech) build robot hardware platforms with integrated brains. Deployment tier represents real-world operation at customer facilities (BMW Spartanburg, GXO Flowery Branch, Mercedes-Benz pilots). The three tiers operate complementarily; understanding which tier a company occupies is essential for evaluating its competitive position and verification posture.
Which companies build foundation models for robotics, and how do they compare?
The brain-provider tier of robotics in 2026 includes several distinct strategic theses. Skild AI pursues cross-platform general-purpose brain deployment. Physical Intelligence (Pi-0; Pi-0.5) emphasizes transformer-based VLA research publications. Covariant specializes in warehouse-automation foundation models. Google DeepMind operates Gemini Robotics and RT-2 across AV and humanoid research. OpenAI Robotics relaunched in May 2026 after a 2021 hiatus. NVIDIA Project GR00T pursues cross-platform humanoid integration aligned with NVIDIA's broader stack. Meta operates research-publication-emphasizing work via FAIR and Reality Labs. The cohort is at research-and-demonstration verification depth; commercial-scale deployment lags behind humanoid OEM commercial deployment substantially.
Why are frontier AI labs (OpenAI, Google DeepMind, Meta) entering robotics in 2025-2026?
The major frontier AI research labs are committing institutionally to robotics in 2025-2026 as a distinct industry cluster. OpenAI formally relaunched its robotics division in May 2026 after a 4-to-5-year hiatus. Google DeepMind operates the Gemini Robotics program as part of its broader model-to-embodiment research. Meta has expanded AI robotics work alongside its Reality Labs portfolio. The cluster is not about any one lab making a robot; it is about foundation-model research lines extending toward physical-world embodiment.