What is physical AI?

What "physical AI" means at DEPLOY

Physical AI is the category of AI systems that operate in the physical world rather than purely in digital environments. The distinction is structural: a chatbot answers text; a physical AI system perceives sensors, decides under uncertainty, and acts in the world.

Per DEPLOY's category framework, physical AI is the umbrella under which four subcategories operate with distinct verification frameworks:

1. Autonomous vehicles + robotaxis. Waymo, Tesla Robotaxi, Zoox, Aurora trucking. Vehicle-form physical AI operating in road networks.
2. Humanoid robots. Figure AI, Apptronik Apollo, Agility Digit, 1X NEO, Tesla Optimus, Boston Dynamics Atlas. Bipedal mobile-manipulation form factor.
3. Autonomous mobile robots (AMRs) + warehouse robotics. Robots operating in structured industrial environments at meaningful scale: warehouse picking, sorting, transport, sub-assembly. Distinct from humanoid form factor.
4. AI-augmented industrial automation. Manufacturing robots, agricultural robotics, construction robotics, drones in inspection and delivery contexts.

Each subcategory has its own verification framework, regulatory environment, deployment cadence, and labor-market interface. The umbrella name (physical AI) is editorially useful for tracking cross-category patterns; per-category analysis requires distinct frameworks.

---

Why physical AI matters as a category

Physical AI is the category where capability claims have to land in operational reality. A digital-only AI system can be evaluated on text benchmarks; physical AI has to perceive, decide, and act under physical-world constraints that produce specific verification surfaces:

• Sensor noise and latency. Camera, lidar, radar, IMU inputs are noisy; perception decisions happen under real-time constraints.
• Mechanical failure modes. Actuators wear; batteries degrade; harmonic drives fail; safety-critical systems have to handle these states.
• Regulatory frameworks. Robotaxis operate under NHTSA Standing General Order + state-level CPUC/equivalent commercial authority; humanoids operate under OSHA workplace safety; drones under FAA.
• Operational envelopes. Each deployment operates against a defined operational design domain (ODD); claims do not generalize across envelopes.
• Per-deployment verification. Capability does not generalize from one demonstration to another; per-deployment evidence anchors verification per-context.

These five surfaces are what DEPLOY's verified-vs-claimed framework operates against in physical AI contexts.

---

The verification framework applied across physical AI

DEPLOY's four canonical frameworks operate across all physical AI subcategories:

• Availability framework (5-tier): consumer-available verified, research-tools verified, enterprise-deployed verified, consumer-promised claimed, engineering-credibility R&D. Worked example: humanoid market.
• Capability framework (4-tier): verified consumer-deployed, verified enterprise-deployed, research-and-demonstration, claimed-future. Worked example: humanoid task capability.
• Safety framework: per-operator verification at cumulative scales; NHTSA Standing General Order + operator publications + third-party actuarial validation surfaces. Worked example: robotaxi safety.
• Value chain framework (brain-provider vs OEM tier distinction): which company builds the platform vs which company builds the AI brain that runs on it. Worked example: humanoid robotics + AV stacks.

Each framework produces per-category insights when applied to a specific physical AI subcategory; cross-category patterns emerge when the four frameworks are applied across all subcategories.

---

Where physical AI sits in the broader AI landscape

Physical AI is the subset of AI where the system's actions affect the physical world. The complement is digital-only AI (chatbots, code-generation, search assistants, content recommendation) where actions affect only digital environments. The distinction matters because verification surfaces differ structurally:

• Digital-only AI can be evaluated on text benchmarks, completion rates, factual accuracy against ground truth.
• Physical AI has to be evaluated against per-deployment operational evidence: how many vehicles assembled, how many totes handled, how many trips completed, how many fatal incidents per million miles.

Both categories use foundation models. Both use reinforcement learning, supervised learning, and similar training methodologies. The differential is in the verification surface, regulatory framework, and operational consequence layer.

---

The cohort by subcategory

Autonomous vehicles + robotaxis (passenger): Waymo (11-metro commercial), Tesla Robotaxi (4-market paid pilot since June 2025), Zoox (2-market free demo pilot), Cruise (wound down). Trucking: Aurora (Dallas-Houston commercial since April 2024), Bot Auto, Kodiak AI, Einride, Plus, Waabi.

Humanoid robots (cohort): 1X NEO (consumer-deployed), Figure 02 + 03 (enterprise-deployed at BMW + Catalyst), Apptronik Apollo (3 Fortune-500 customer pilots), Agility Digit (GXO 100K-tote anchor), Tesla Optimus (consumer-promised, internal-only), Boston Dynamics Atlas (engineering-credibility R&D), Unitree G1 + R1 (research-tools), Sanctuary AI, Mentee Robotics, Engineered Arts, PAL Robotics.

Autonomous mobile robots + warehouse robotics: Symbotic (Walmart $200M deal January 2025, $22.7B backlog, 42 distribution centers), Locus Robotics, Geek+, AutoStore, others operating at meaningful enterprise deployment scale.

AI-augmented industrial automation + drones: extends to manufacturing robots, drones in delivery/inspection contexts, agricultural robotics (orchard, dairy, field), construction robotics. Category boundary expands as deployment evidence accumulates per vertical.

---

Bottom line

Physical AI is the category of AI systems that operate in the physical world under verification surfaces that digital-only AI does not face (sensor noise, mechanical failure modes, regulatory frameworks, per-deployment operational evidence). DEPLOY tracks physical AI across four subcategories with distinct verification frameworks; the umbrella name is editorially useful for cross-category pattern tracking. Per-category analysis requires distinct frameworks because the operational envelopes, regulatory frameworks, and labor-market interfaces differ structurally across robotaxis, humanoids, AMRs, and industrial automation.

For per-subcategory deep-dives, see where Waymo operates for AV/robotaxi, what can humanoid robots actually do today for humanoid capability, what is a foundation model for robotics for the AI-brain layer, and brain-provider vs OEM platform tier distinction for the value chain framework. For methodology canonical references applicable to physical AI category umbrella: the 4-way autonomy-boundary taxonomy + verified-vs-claimed at within-entity granularity + captive vs third-party brain providers + the 9-tier source-quality rubric.