Duck Host

The robot already knows your home. Now it knows who lives there.

House Twin maps a home from one iPhone walk, and a Microduck robot patrols it by voice. Duck Host adds Qloo: before the duck acts, it asks what each person in the room will actually like, and shows the score.

"Find something we can all watch tonight." (Rosa, 78)
  1. heard Omni turns speech into text
  2. plan patrol → living_room
  3. tool pick_for(living_room, [Rosa, Ana, Leo], film)
  4. Qloo 7 films × 3 people, each scored
  5. pick the best lowest score
Tonight in the living room Paddington 2
The sample home from House Twin (samples/synth-multi), drawn from its house.json. Scores on this page are illustrative.

One room, three tastes

Each person sets a few favourites once. When someone asks for a film, Qloo scores every option for every person, and the duck picks the one whose lowest score is highest. Change a favourite and watch the pick move.

Why this only works with QlooA language model can name family films. It can't measure how much Rosa, Ana and Leo will each like one, so it can't show the bars, and it can't prove the pick is fair.

The dashed row is a generic "family movie night" pick, scored the same way for comparison. All numbers are illustrative until the live Qloo scoring test passes.

At the door: what's worth doing nearby

In the hallway, Rosa asks where to have lunch. A map app sorts by distance. Qloo ranks the places near home by how well they fit her taste. For someone who doesn't go out much, the same ranking picks what to order in.

    Qloo gets one rounded neighbourhood point and Rosa's favourites as taste IDs. Place names are invented for this page.

    Every room gets a mood

    The organisers' own example is room lighting that follows mood and taste. Qloo's ambience tags for Rosa's favourites set the light colour and the music in the room the duck is visiting.

    Kitchen
    Warm white

    On its own, a model can roughly guess a mood from favourites, so this is decoration on top of the scores, not the proof.

    The check-in round brings one new thing

    House Twin already does check-in rounds for someone living alone. Now each round ends with one suggestion, one step beyond what Rosa already loves, with the reason. She answers by voice, in her own language, and her profile changes.

      Rosa's taste profile

      Pedro Infante filmsColumboBuena Vista Social Club

      Only taste IDs are stored, never the recording. "Me gusta" adds the title as a new signal for Qloo.

      How it works

      The agent gets two new tools next to patrol and ask_person. The Qloo key and every call live on the House Twin server. The browser and the robot never see the key.

      House Twin server: the key, profiles and floor plan stay here Voice or text"Busca algo para ver"any of 6 languages Agent (Ultra or Claude)reads house.jsonpicks rooms and tools patrol · ask_person pick_for(room, people)new go_out(kind)new Qloo APItaste IDs in,scores out Duck walks+ result cardon the hosted page

      What Qloo receives

      • Taste IDs for films, shows, artists and places
      • One rounded neighbourhood point, for nearby places
      • An age group, if the household sets one

      What never leaves the house server

      • Names, voices and recordings
      • Camera frames and depth scans
      • The floor plan and device IDs

      What Qloo adds to House Twin

      House Twin existed before Sep 30. This is the list of what is being added for the Qloo Agentic Hackathon. The baseline is the tag baseline-nebius in the repo.

      • A server-side Qloo client, with a cache, and the key kept out of the repo and the browser.
      • Two agent tools, pick_for and go_out, alongside patrol and ask_person in agent/language.py.
      • A household file of taste profiles next to house.json, holding Qloo IDs only.
      • A hosted page judges can use with no robot: pick a household, give a command, watch the recorded walk and the scored result.
      • A test: the same command for three households must give different picks.
      • A README section with before and after, and every Qloo call logged like the Nemotron calls.