The problem
In one week, three ways to give a factory robot its brain showed up side by side.
On 10 November Google DeepMind released Gemini Robotics 2.5, a vision-language-action model it tested on more than 200 tasks. The headline result came from Apptronik’s Apollo 2 humanoids on active Mercedes-Benz lines in Kecskemét and Sindelfingen: kitting metal brackets from bulk bins into sequencing containers with a 96% success rate. The model is rolling out to DeepMind’s robotics partners. Humanoids Daily notes there is no announced API access or pricing.
Xiaomi took the open route with Xiaomi-Robotics-1, an Apache 2.0 model pre-trained on more than 100,000 hours of real-world manipulation trajectories, with a 5B checkpoint on Hugging Face and the post-training code on GitHub. Separately, on 3 November Xiaomi showed its humanoids on an eight-hour sorting task at its car factory; Electrek does not say those robots were running Xiaomi-Robotics-1.
Embodied AI, a 25-person EPFL spin-out, took a third route. On 6 November it unveiled AIR, a two-armed robot on a wheeled base that runs a stack of task-specific models instead of one general robot model. A remote operator drives it from the first shift, every shift is recorded, and the company says the robot takes the task over on its own in about a month.
For a plant or a robot maker choosing how to buy robot intelligence, these look like three competing bets. I read them as three answers to the same question, and the answer they share matters more than where they differ.
What I concluded
In all three, the job on the line was learned on the line. Apptronik’s CEO says the Gemini models “improve dramatically in just a few weeks of post-training” before carrying over to Mercedes-Benz production. Xiaomi says its model learns a new complex task “from just a few hours of demonstrations per task”. Embodied AI’s whole pitch is a month of teleoperated shifts that become training data.
So the base model is the part you can rent. The asset that builds up on a specific line is everything else: the demonstrations, the failure clips, the intervention logs, and the acceptance tests for that station. My recommendation: pick the base model on capability and access, and negotiate as if the line data is the thing you are buying. Who keeps it, in what format, and whether you can train a different model on it next year.
Constraints
Why it is hard to decide on public information alone:
None of the headline success rates comes with a denominator. DeepMind’s 96% does not say how many picks it covers, how long the pilot ran, or how cycle time compared with a person; Humanoids Daily points this out. Embodied AI’s 95% is an average across paid pilots, with undisclosed tasks, definition of success and duration. Xiaomi did not publish part counts, the success rate over its eight hours, or how many times a human stepped in, as Electrek notes, although it did publish rates for earlier stations (90.2% in March, 98% by July).
Base models move fast. Gemini Robotics 2 shipped in July and 2.5 in November. A line that depends on a vendor’s model inherits that release schedule, good and bad.
Access is uneven. DeepMind says it is rolling out to robotics partners; Humanoids Daily notes there is still no announced public API or pricing. The open model is free to download, but the integration and the safety case become yours. The on-site stack comes bundled with a vendor and its operators.
Nobody published data terms. None of the three sources says who owns the demonstrations and logs collected on a customer’s line, or whether a customer can take them elsewhere. Embodied AI says “what one robot learns can be shared with others”, which is good for the fleet and is exactly the clause a buyer should read closely.
The options, and what I traded away
The decision: a plant (or a robot maker selling to one) has to pick where the robot’s intelligence comes from for a set of manual stations.
| Option | What it buys | What it costs | Why I did / did not pick it |
|---|---|---|---|
| License a frontier VLA through a partner robot (Gemini Robotics 2.5 on Apollo 2) | Broadest skill set on public evidence: 200+ tasks, and DeepMind reports 30% to 100% higher success than previous baselines, depending on task type; a published factory pilot | Rolling out to robotics partners (DeepMind); no announced public API or pricing (Humanoids Daily); model updates on the vendor’s schedule; data terms not public | Would pick it when the stations are varied and dexterous, but only with data rights written down |
| Start from open weights (Xiaomi-Robotics-1) | Apache 2.0, fine-tune on your own demonstrations, run on your own hardware, no access gate | You own integration, evaluation and the safety case; the headline benchmarks are the developer’s own | Would pick it for a team that already has robotics engineers and wants the data loop in-house |
| Buy a vendor’s narrow stack learned on site (Embodied AI AIR) | Productive from shift one via teleoperation; a policy fitted to your exact station within about a month | Operator labour during learning and for exceptions after; the learning loop lives with the vendor; no price published yet | Would pick it for a few high-variation stations where time to first output matters more than owning the model |
| Wait for the models to settle | No integration work spent on a model that may be replaced in four months | Nothing collected; the stations keep running by hand and no line data accumulates | Would not pick it. Recording demonstrations and interventions is useful under any of the three paths |
Evidence
The frontier path is strong on breadth. DeepMind reports the 2.5 model succeeding 30% to 100% more often than earlier baselines depending on task type, and finishing 20% to 30% faster. Humanoids Daily lists per-task jumps against Gemini Robotics 2, such as packing pastries with tongs going from 5% to 100% and folding pants from 0% to 40%, without trial counts. At Mercedes-Benz the work was split: an embodied-reasoning model tracked the workflow and the action model moved the hands. Even here, the CEO’s own description of the path to the line is “a few weeks of post-training” on Apollo 2.
The open path makes the same point from the other side. Xiaomi’s README promises out-of-the-box mobile manipulation after post-training on more than 10,000 hours of cross-embodiment data, and then “a few hours of demonstrations per task” for new jobs. The weights are free; the hours of demonstrations at your station are not.
The on-site path makes the data loop the product. Embodied AI’s site says its compact policies are orchestrated by its own vision-language model running on the robot, and that “models trained in simulation break when reality shifts”. Its software already runs in a paid two-arm cell and, per GGBa, hit 98% reliability on a consumer-electronics maker’s cable-handling task against a 90% target. After handover, operators stay available for situations the robot has not seen.
Read together: the three differ on who trains the base model, and agree that someone has to collect task data at the station. Whoever holds that data can switch base models later. Whoever does not has to collect it again.
Counter-arguments worth taking seriously. “Frontier models will get good enough that per-line post-training disappears.” Possibly, and about four months between Gemini Robotics 2 and 2.5 is an argument for it; the Mercedes pilot still involved post-training, so I would not plan on it yet. “Data rights are a procurement detail.” They are a detail until you want to change vendors and find your month of demonstrations sits in someone else’s training set. “Open weights are only for labs.” Xiaomi ships deployment code and targets consumer GPUs; the real barrier is evaluation and safety work, not access.
Where this stops being true
The argument is about manual, high-variation stations where a robot has to be taught the specific job. It weakens for a single repetitive task where a fixed automation cell is cheaper than any learned policy; once a base model reliably does a station zero-shot, with published denominators, so there is little line data worth owning; and where a vendor contract already gives the customer full copies of demonstrations, logs and evaluation sets. I am not claiming any of the three approaches is wrong, or ranking their models. The public figures don’t support a ranking.
What I would do differently
If I were running a plant pilot this quarter:
Start recording now, whichever path I pick: demonstrations, failure clips, every human intervention with a timestamp and reason.
Write the station’s acceptance test myself: picks per hour, success over a full shift, interventions per hour. Run every candidate against it, whatever its brochure says.
Put three clauses in the contract: I get a copy of all data collected on my line in a documented format; I can use it to train another model; and data shared across the vendor’s fleet is opt-in.
Ask each vendor the same three numbers for its headline rate: how many attempts, over how long, and how many interventions.
Choose the base model last. It is the part most likely to change, and the part easiest to swap if I hold the data.