Matic put its privacy boundary at the wake word, and published where it sits

Matic shipped Cues on August 13, an opt-in software update that lets an owner of its robot vacuum point at a spill and say "clean this," send the machine to a named room, or tell it to follow them ^[1][2]. The notable part is not the pointing; it is that the company stated where the compute boundary falls, because that boundary is the real privacy specification and most home-robot vendors never publish theirs.

On the robot side, Matic says three models run locally: wake phrase, voice direction, and gesture recognition ^[9]. The launch page describes the gesture path as detecting a pointing hand, estimating its direction in 3D, and using that direction to locate the area to clean ^[5]. Camera data used for gestures stays on the device and is discarded, according to the company ^[10]. Microphones are off by default and can be turned off again in settings, per Matic's privacy documentation ^[8]. The microphone array is specified to pick up a command from up to 10 meters ^[7].

The cloud side starts one step later. Once the robot detects "Hey Matic" on-device and plays a chime, the command that follows goes over an encrypted connection to Google's Gemini API for parsing ^[11]. Matic says the request carries no user information and is not stored ^[12]. So the line sits precisely at wake detection: perception is local, language understanding is remote ^[15].

The reason for that line is coverage. Matic claims support for more than 70 languages through the voice interface ^[3], and the same command set spans room vacuuming and mopping, carpet edges, running a schedule, and come-here behavior ^[6]. The company says it plans a smaller on-device parser, but that it will probably handle fewer commands and may initially support English only ^[13]. That is a useful admission of price: local parsing costs you vocabulary and languages, and the vendor has quantified the direction of the trade even if not the magnitude.

For anyone specifying connected hardware, the transferable part is the question format. "Is it processed locally" is unanswerable for a pipeline with four stages. "Which stage crosses the network, in what form, to whose model, and what confirms the crossing to the user" is answerable, and Matic has answered it: the audible chime is the signal that the outbound stage has begun, which makes consent and wake confirmation product controls rather than policy language ^[16]. Ask for that stage-by-stage map before the demo, and note when a vendor cannot produce one.

Two things remain unestablished. The publisher notes that launch materials and a single hands-on report do not demonstrate reliability across varied rooms, accents, background noise, pointing angles, and unusual floor layouts ^[14]. And the no-retention claim is Matic's own statement about its requests ^[12]; the source material does not say whether it is backed by a contractual term with Google.

Worth watching: whether the on-device parser ships and what command coverage it actually lands on ^[13], whether false wakes push audio to the cloud in normal household noise, and whether Matic keeps documenting the boundary as the feature set grows. The Verge described Cues as a free update ^[4], which means the fleet moves to this architecture without anyone re-reading a purchase decision.