What Is Physical AI?

Start with a physical decision

For this database, physical AI describes learned perception, reasoning or action used in a system that interacts with the physical world. This is a working editorial scope, not a certification. Google DeepMind’s Gemini Robotics family illustrates the action-model layer, while a deployed robot also needs hardware, control and an operating environment. The useful question is which decisions the learned system makes and which other components constrain it.

Separate the objects being evaluated

A company, model and robot are different objects. A developer may publish a model without selling a robot, while a robot supplier may combine several software components. Begin a shortlist with the exact system and version that could perform the intended task. Then trace its developer, software, training or evaluation evidence and commercial access. Brand familiarity cannot fill a missing link in that chain.

Describe one complete task

Write the task as a beginning, action and observable end: for example, moving a tote from one station to another under stated conditions. Record object variation, people nearby, workspace changes and recovery requirements. GXO’s Digit announcement is useful because it describes a concrete logistics context. It is evidence about that operation, rather than a claim that the same robot can perform every warehouse job.

Keep research, demonstration and operation separate

A simulation score, edited demonstration and customer-operated workflow answer different questions. LIBERO evaluates specified simulated manipulation tasks; a deployment source concerns an installation and workflow. Neither should erase the other’s limits. Look for evidence closest to the decision being made, and keep unanswered fields visible. A missing intervention rate is an unanswered operating question, not evidence that intervention never occurs.

Use the graph to decide what to investigate

Follow a company to its product, model, evaluation and deployment records. A broken evidentiary link is a research question: which model revision is installed, what task was tested, or who confirms operation? For an initial assessment, identify the proposed task, required human support, acceptance test and buyer access path. These details make the term physical AI useful beyond a broad market label.

Additional disclosed fields

Action-model example

Gemini Robotics converts visual and language inputs into robot actions

Company claimReviewed 2026-09-28

Sources 1

Gemini Robotics model family

Google DeepMind · Publication date not disclosed · Source accessed 2026-09-28

Physical-system example

Spot is an inspection robot with an associated software ecosystem

Primary specReviewed 2026-09-28

Sources 1

Boston Dynamics — official portfolio and company information

Boston Dynamics · Publication date not disclosed · Source accessed 2026-09-28

Official page read during live research; confirms identity and listed offering. Performance language remains a company claim.

Operational evidence example

GXO publishes a Digit commercial agreement and defined warehouse workflow

Primary releaseReviewed 2026-09-28

Sources 1

GXO multi-year agreement with Agility Robotics

GXO · Published 2024-06-27 · Source accessed 2026-09-28

Customer describes commercial live-warehouse operation after a pilot. Does not identify current Digit 5 hardware.

Relationships & deployments

Related reading is an editorial crosslink. Sourced connections describe relationships reported in the cited material. A link to a versioned profile does not establish compatibility with that version unless the connection note explicitly identifies it.

Record history & verified changes

A research review records when we checked a source. It does not mark a product launch or a new deployment.

Initial reviewed record. No subsequent field change has been recorded.