Not a faster drink machine.
A bartender that lifts its head to look at you.

The market has solved making the drink. We are solving something else: making the drink worth watching — and in this version the arm's motion is learned, not recorded.

Their AI is in recommendation, not in motion

Makr Shakr 240 drinks/hour, Cecilia.ai $45,000, Yanu in a sealed pod: optimised for throughput and consistency, and they hit it. The installation is built around the machine — glasses in a jig, bottles on a rail, motion along a fixed path.

Robots that talk exist; arms that move exist; nothing puts both in one body, and every arm replays a recording. Move the glass 5 cm and the same trajectory misses.

One service is nine stages; gaze appears twice

A bar has never only sold liquid. People sit at a bar because there is a person, a performance, an exchange. Existing systems optimise speed for places with a queue; we optimise the 40 seconds you are watched — craft bars, events, showrooms.

The head is the interface

Two bartenders on the staff board — one body

Barley

Head bartender · the warm shift

Warm, quick and a little silly. One or two sentences, already reaching for the glass while talking, and the line while fetching ties their drink to them.

Teases the drink, never the customer. Never says a price.

Cyborg

Bartender on duty · the dry shift

Listens to the small talk — just will not follow you up into the clouds. When the customer turns wistful it pulls the conversation back down to the floor: "that is not a philosophical question."

Dry, practical, paying attention the whole time. The opposite temperature, the same body.

A bartender is a text file here. Swap the file and you have swapped the bartender; the head wears the outfit that goes with it. That is how one machine fits a craft bar and a trade-show stand without rebuilding anything.

We fixed the data, not the code

The right-arm model scored 0/14 on 09-08. Diagnosis: cup-position coverage in the 40 demos was 16× unbalanced (16 episodes in the densest band, 1 in the sparsest); the model had learned tempo, not vision — closure timing correlated with cup position at only r = 0.16–0.26.

Not a replay: it changes its mind before knocking the cup over

That is what a closed loop looks like: miss → see → correct. A flawless run proves nothing; a correction proves the policy is reading the image.

How a small model "chooses": put the choice in front of the model

The training set has one cup per frame and one sentence. With two cups on the table it reaches between them (three times on 09-11); adding "yellow" to the instruction perturbs the trajectory but not the target. SmolVLA has 450 M parameters on a 7.4 GB board; teaching it to select by word means re-recording and a night of training, with no precedent that it would learn.

What the night run found · 2026-09-12

The rule changed from "paint out the cup of colour X" to "keep one, paint out everything else". A second distractor on the table should not need a config change.

On a 25-frame evaluation set the colour thresholds we had measured in daylight on 09-11 detected 1 of 25. At night the table and the cups both fall to saturation 10–30 and colour stops being a separable feature at all. Rebuilt around the table region plus brightness relative to the table: 15 of 25.

The gripper cannot be excluded by colour either — the cream cup and the orange bracket share one hue band (H 8–30), so excluding orange kills the cup. It is excluded by position: both arms always enter from the bottom corners.

And the "empty table" plate we meant to paste over distractors turned out not to be empty — it had a green cup in it. Pasting it would have inserted a cup that is not there. Median fill instead: no plate, no shape, adapts to the light.

measured25 real head frames, day and night: the 09-11 colour rule detects 1/25, the rebuilt detector 15/25. Two new channels shipped — --mask-keep (keep one cup, paint out the rest) and --mask-box (paint out a rectangle). Runs in the pipeline with per-order switching, 3 ms/frame. Eight exploratory trials on the left-arm model went for the blue cup whichever cup was masked — that model carries its own bias and is not a valid testbed. not measuredselection accuracy — needs a preregistered 5 + 5 on the right-arm model (mask blue → grasp yellow, mask yellow → grasp blue) with the criterion "closure pan on the target side". The A/B on 2026-09-12 does not count: there was no target cup in the grasp area, so both rounds were held=0 by construction and the 2% safety-layer difference is single-trial noise. It needs a target and a distractor on the table, and someone on site to place them.

Operated from 7,000 km away — a feature, not an apology

The team is in China, the robot in Finland. Every trial on 2026-09-11 — launch, judging, logging — was run remotely: the panel starts the job, the person on site places the cup and presses confirm, results are archived automatically.

Three sites, one robot. The arm is in Helsinki; the people who launch the runs, judge them and write them up are 7,000 km away in Hangzhou and Taiwan. Nothing on this deck needed anyone to fly.

The safety layer, and the rules for counting

Why bother. Not modesty — this is what lets us say the interesting things and be believed. The three numbers on slide 5 are what the discipline produces: 6/6, 3/8, 1/5, kept apart rather than averaged into one figure that would flatter us and tell you nothing.

It runs on a palm-sized board

Two arms, a moving head, three cameras, a mobile base. The board cannot hold a 7-billion-parameter model — so we run the 450 M SmolVLA and hand problems like "which cup" to 3 ms of classical vision (slide 6). The constraint forces a smarter design, not a bigger model.

Honest limits, each with its next step

Four weeks after selection, this is what we will have

A promise, not a wish list. Every line below carries the condition that decides whether it passed, and every result lands in the public report space with its date and its k/n — including the ones that fail. You will not have to take our word for any of it.

1. Week 1 · Technical
   Record the colour-instruction data

   50 episodes, half yellow and half blue, the two positions swapped, trained together with the single-cup set.Done when: the dataset is public and the coverage CV across the five cup-position bands is reported.

2. Week 2 · Technical
   One preregistered round, whatever it says

   Criterion, trial count and void rule fixed before the run, then judged under the slide-8 rules.Done when: the k/n is published — 6/6 or 3/10, it goes on the same page.

3. Week 3 · Product
   Drinks that actually have to be made

   Pouring moves inside the pipeline: a POUR state and a pour() effector. Today coke, fanta, sprite and water are four names for one physical skill; this ends that.Done when: at least two drinks of two or more ingredients, poured in order to a recipe, run end to end on video.

4. Week 4 · Product
   Not once — every time

   Ten consecutive runs of the same configuration with no hand in the loop, and the hand-over direction taken from the measured face bearing rather than a fixed coordinate.Done when: the ten-run k/n is published under the same rules.

5. Week 4 · Business
   In front of a real crowd

   The 10-step arrival procedure at a new venue, two hours from boxes to a run; one cocktail event or bar week agreed and attended; feedback collected on the night.Done when: the event happened, the feedback is written up and published whatever it says, and a named venue has stated in writing what would put the robot on their floor.