Apptronik's Apollo 2 Humanoids Start Kitting Work at Mercedes-Benz

Two humanoid robots now spend part of their shift reaching into blue plastic bins on an active Mercedes-Benz production line, pulling out metal brackets, and dropping each one into a sequencing tray in the exact order an assembly worker downstream will need it. Google DeepMind said on Tuesday that the robots, built by Apptronik and running its new Gemini Robotics 2.5 model, complete that kitting task correctly 96 percent of the time at Mercedes-Benz plants in Kecskemét, Hungary, and Sindelfingen, Germany.

The deployment is a proof of concept, not a production rollout, and DeepMind has not disclosed how many picks that success rate covers, how long the pilot has run, or how its cycle time compares with a human doing the same sorting job.

What a 96 Percent Pick Rate Actually Covers

The task is called automotive kitting: taking individual parts out of bulk storage and arranging them into organized containers in the sequence the vehicle line will consume them, so a technician never has to stop and search for the right bracket. DeepMind says Gemini Robotics 2.5 handles the gap correction this requires, adjusting grip and placement when a bin runs low or a part sits at an awkward angle, which is the part of kitting that has historically kept it a manual job even in otherwise automated plants.

The company frames the Mercedes-Benz line as one data point inside a broader claim: that Gemini Robotics 2.5, tested across more than 200 distinct tasks from serving food with tongs to pipetting liquid, beats its predecessor's success rate by 30 percent to 100 percent depending on task type, while finishing tasks 20 percent to 30 percent faster. DeepMind also says reasoning-based safety guidance at inference time raised safe task completion by 33 percent around obstacles and by 43 percent when a person is nearby, and that a new internal safety benchmark, ASIMOV-Jailbreak, scored 99.5 percent against the prior model's 82 percent. None of those figures have independent verification yet; they are DeepMind's own measurements, published alongside the model rather than audited by a third party.

A 96 percent success rate on sorted brackets is a statement about a narrow, controlled task, not a proof that a robot can handle the variation an actual assembly line still routes to a human.

Why Kitting Was the Entry Point, Not Welding

Kitting is a deliberate choice of starting task, and it says something about how both companies are managing risk. Unlike welding, torque-critical fastening, or anything that touches a vehicle's structural integrity, a misplaced bracket in a sequencing tray is a recoverable error: a line worker downstream notices a gap and compensates, rather than a defect reaching a finished car. That tolerance for mistakes is exactly what makes kitting a reasonable first humanoid task on a line that still has to hit quality targets every shift, and it is also why the 96 percent figure should be read as a floor for an easy case rather than a ceiling for what the technology can do. Industrial automation has handled bin-picking with fixed robotic arms for years, but almost always with heavy fixturing, known part geometry, and a cage around the cell; what is different here is a general-purpose, mobile, two-armed machine doing it without that scaffolding, in a human-occupied aisle, next to a worker who is not wearing safety glasses because of the robot.

Mercedes-Benz is not alone in treating a car plant as the proving ground for humanoids. Figure AI has run a comparable dexterous-sorting deployment inside BMW's Spartanburg, South Carolina plant, and Hyundai has put Boston Dynamics' Atlas through structural and mobility testing at its own facilities. The pattern across all three is the same: automakers have the capital, the controlled floor space, and the existing safety culture to host a pilot, and they get early access to a capability they may eventually use themselves. For a procurement team outside the automotive sector watching from the sidelines, the Mercedes-Benz pilot is best read as a signal about how fast the underlying foundation models are improving, not as evidence that a similar robot is ready to walk onto a warehouse floor with a different part mix and no automaker-scale integration budget behind it.

Argon's Simulation Loop Is Doing Work the Video Doesn't Show

The more interesting mechanism behind the Mercedes-Benz numbers is not the robot arm, it is the model that trained it. Gemini 4 Argon, a separate frontier model Google released in October, generates thousands of varied simulated practice runs from a single real-world camera scene and builds a digital twin of it, then combines those simulated reps with real demonstrations so the robot generalizes to new objects and layouts instead of memorizing one bin configuration. Argon also reviews footage of the robot's failures, proposes a cause, and tests a fix in simulation before it goes back onto the real machine.

That loop, simulate, fail, diagnose, retest, is what let Apptronik's chief executive Jeff Cardenas say the models improved substantially within a few weeks of post-training and carried that improvement over to the Mercedes-Benz floor with little performance drop. DeepMind has not said how much of the four-month gain between Gemini Robotics 2 and 2.5 came from Argon specifically versus additional real-robot data collection or partner-side post-training, which matters for any buyer trying to estimate how repeatable this is on a different line with a different part mix.

Apptronik Supplies the Hands, DeepMind Supplies the Judgment

Apptronik, the Austin, Texas-based humanoid robot maker, builds the hardware DeepMind's software runs on. Apollo 2 is its second-generation platform: a modular design that ships in a bipedal configuration for spaces built around people and a wheeled-base configuration for high-throughput work, built on what the company calls a patented actuator platform aimed at keeping energy use, maintainability, and manufacturability in check rather than chasing peak specs. On the Mercedes-Benz line, the robots work with a five-finger, 22-degree-of-freedom hand called SharpaWave, made by Sharpa, while a separate Gemini Robotics-ER model handles the higher-level job of tracking the multi-step workflow and Gemini Robotics 2.5 itself controls the arms and hands.

That division of labor, one company's body, another's foundation model, a third's hand, is becoming the default structure for humanoid deployments rather than the exception, and it is also why DeepMind's access policy matters as much as its benchmark numbers: Gemini Robotics 2.5 is currently limited to DeepMind's existing robotics partners, with no public API, no pricing, and no new on-device or embedded-reasoning model announced alongside it. A procurement team cannot simply buy this capability; it has to already be inside DeepMind's partner circle, which today means hardware makers like Apptronik rather than end customers like Mercedes-Benz directly.

The Distance Between a Demo and a Payroll Line

Nothing in Tuesday's announcement changes the basic fact that this is still a pilot running alongside, not instead of, existing staffing on two plant floors. The headline number, 96 percent, describes a single task picked specifically because it is well suited to a camera-and-gripper system: discrete parts, a known bin, a known tray. It says little about the far larger set of assembly-line jobs that involve judgment calls, awkward reaches, or parts that do not come pre-sorted into bulk bins at all. The gap between a demo that works on video and a system a plant manager would actually staff a shift around is still measured in the details DeepMind left out here: pick count, pilot duration, and a direct cycle-time comparison against the technician the robot is meant to free up for harder work.

For companies weighing whether to follow Mercedes-Benz's lead, the more useful signal is not the success-rate headline but the model behind it: a frontier lab is now explicitly building a simulation-and-failure-analysis pipeline as a product feature, not a research side project, which suggests the next round of humanoid pilots will be judged less on raw dexterity demos and more on how fast a given model can close the gap between a lab benchmark and a specific customer's floor.

The access terms attached to Gemini Robotics 2.5 point to where the real decision point sits for buyers. Because the model ships only to DeepMind's existing robotics partners, with no API and no published pricing, the commercial question for a procurement team is not whether to license this specific model but which hardware partner to bet on, since that partner's access to the underlying foundation model is now effectively bundled with the robot. That makes the vendor-selection questions sharper than they were a year ago: how many foundation-model partnerships does a given humanoid maker actually hold, how portable is its hardware across more than one partner's software stack, and what happens to a fleet's capability roadmap if that partnership ends. A buyer evaluating Apptronik, Figure AI, Agility Robotics, or any other humanoid maker today is, in practice, evaluating a software relationship as much as a piece of hardware.

By the end of the shift at Sindelfingen, the two Apollo 2 units are still working side by side with human technicians a few meters away, reaching into the same blue bins, under the same factory lights, their hands now doing the one job nobody on that line ever wanted to do by hand in the first place.

This analysis is for general information purposes only and does not constitute investment, legal, or professional advice. It synthesizes Google DeepMind's and Apptronik's own public statements and published benchmark disclosures as of publication.