Physical AI Buyer Checklist

Define the job and its exceptions

Document the input objects, starting state, required output, shift pattern and physical constraints. Include common exceptions: damaged packaging, blocked routes, changed object positions and recovery after interruption where relevant. Choose an observable completion rule and identify who decides whether it was met. GXO’s tote-transfer account is a useful model of a narrow task boundary; general warehouse automation is too broad for an acceptance test.

Freeze the proposed configuration

Record robot model, hardware revision, end-effector, sensors, software and model versions, power arrangements and connected systems. Ask the vendor to identify which specifications are final and which are preliminary. Then map every important claim to evidence for that configuration. Do not borrow a payload from one robot, autonomy from another release and deployment history from an older generation to create an imaginary combined product.

Budget for integration and people

Assign responsibility for site preparation, interfaces, calibration, task setup, remote assistance, maintenance and exception handling. Record operator time as well as robot task time. Proxie’s fleet tooling and NEO’s described expert assistance illustrate that software and human services can be part of the operational system. Ask what happens when connectivity, an integration or the primary task fails.

Design the acceptance trial before negotiating outcomes

Specify a representative workload, variation, observation period, failure categories and intervention logging. Include recovery, restart and degraded conditions that are relevant to the site. Require raw counts and their denominators alongside percentages. A customer-reported peak rate, such as the one in DHL’s account, still needs workload context before it can become an acceptance threshold for another facility.

Make the purchase conditional on evidence

Request a dated commercial offer, support scope, delivery configuration and agreed test results. Record unresolved questions explicitly and assign the next verification action. A useful decision may be proceed to a scoped trial, defer pending evidence, or reject for a task mismatch. None requires a universal winner. Maintain the evidence record after installation so upgrades and changing operations trigger a fresh review.

Additional disclosed fields

Bounded workflow example

GXO describes tote transfer between existing automation and conveyors

Customer confirmedReviewed 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.

Human-support example

NEO describes scheduled remote expert assistance for unfamiliar tasks

Company claimReviewed 2026-09-28

Sources 1

NEO order and hardware

1X Technologies · Publication date not disclosed · Source accessed 2026-09-28

Fleet-system example

Proxie is offered with associated fleet-management software

Primary specReviewed 2026-09-28

Sources 1

Collaborative Robotics — official portfolio and company information

Collaborative Robotics · 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.

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.