Nobody writes poems about kitting. On a car assembly line, metal brackets sit in bulk bins. Someone has to lift them out one at a time and set them in sequencing trays, so that the next worker down the line finds each part where it should be. According to Google DeepMind, that job was recently done at Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany, by a humanoid robot called Apollo 2, built by Apptronik and steered by Google’s artificial intelligence.
It got the job right 96 times in every hundred.
That figure was the centrepiece of the launch of Gemini Robotics 2.5 on November 10, which Ryan Whitwam reported for Ars Technica. DeepMind’s announcement, signed by robotics head Carolina Parada, calls it the company’s “most capable vision-language-action (VLA) model to date.” Like every rising power, it arrives with its achievements counted in percentages.
The anatomy of a hand #
The system has two layers. The first, Gemini Embodied Reasoning, is a vision-language model: it takes a human instruction, watches live video and works out how the task could be done. The second, the vision-language-action model now numbered 2.5, turns that plan into movement, the turning of an actuator or the closing of a finger. In the Mercedes pilot, DeepMind says, the reasoning model kept track of the workflow while the action model ran the hands and arms.
In the laboratory it was tested on more than 200 tasks: serving food with tongs, drawing liquid with a pipette, tying a bow. The bow matters. Ribbon, like anything that bends and slumps, is notoriously hard for machines to handle, and Apollo 2 tied one with its five-fingered hand, which DeepMind identifies as a SharpaWave hand with 22 degrees of freedom. Depending on the task, the new model succeeds 30 to 100 percent more often than its predecessor and finishes 20 to 30 percent faster. DeepMind’s own charts are blunter. Closing a toolbox lid went from 0 percent to 93. Packing pastries with tongs went from 5 to 100. Untying ribbon bows went from 0 to 70. Hanging a cardigan went from 0 to 62.
The last four percent #
The factory is not the laboratory. A person doing repetitive kitting gets it right almost every time, and for a carmaker 96 percent is probably still too low. Kanishka Rao, a DeepMind director, said that closing the last 4 percent would be a “steep road.”
For decades industry has had robots that do a handful of things perfectly. They are programmed by hand, bolted in place and blind to everything outside their task. DeepMind is after something more ambitious and more fragile: a general machine that can be put to almost any job. Parada described the bargain plainly: “We can be very good at a particular application at the cost of generalization. Or we can just push on the generalization axis and maybe not get to 99.9 in every application.”
Apptronik’s chief executive, Jeff Cardenas, is quoted in DeepMind’s post as saying the models “improve dramatically in just a few weeks of post-training” and carried over to the Mercedes lines “with very little drop-off.” The post presents the work as a proof of concept with one industrial partner. It is not a deployment.
A rehearsal in simulation #
Chatbots were fed more or less the whole open internet. Robots have no such library of physical experience. The best training data comes from teleoperation, people steering machines by remote control, and that is slow to gather. So DeepMind used Gemini 4 Argon, its frontier model, as a simulation engine. From real camera footage of a scene, Argon generated thousands of variations of practice runs, digital twins in which the robot could rehearse. The team says this helped the model adapt to new tasks and settings faster. The synthetic runs were then mixed with real demonstrations to keep the model tied to reality.
Argon was also used to study failures from video and suggest fixes, tested in simulation and on real robots. That analysis does not happen in real time, because large models are slow to reason and a robot cannot stop to think in the middle of a movement. Even the action model runs on cloud servers, not on the machine itself, so the gains in speed still come with latency. The demonstration video was played at double speed to make the robot look more human.
Learning to refuse #
A machine that works next to people has to know when to stop. DeepMind says Gemini Robotics 2.5 reasons as it acts about hazards, halting entirely or waiting when something blocks its path. With this guidance, it safely completed tasks 33 percent more often when obstacles were present and 43 percent more often when people were in the way.
There is still no agreed way to test physical AI for safety. DeepMind’s proposal is ASIMOV-Jailbreak, an open-source set of prompts meant to trick a robot into doing something dangerous. Ars Technica puts the number of prompts at 1,500. DeepMind’s post speaks only of “hundreds of prompts that disguise unsafe requests across dozens of scenes.” In one example, a robot is asked to pack items in a bag and close it with a fastener too fiddly for its fingers. The correct response is to decline. Trying anyway, and closing the bag badly, counts as the unsafe outcome.
The scores do not agree either. Ars Technica reports that putting Argon in the loop yields a 98.8 percent jailbreak safety level, against 73.5 percent for Gemini Robotics ER 2 alone. DeepMind’s post gives Gemini 4 Argon 99.5 percent and the earlier baseline, Gemini Robotics ER 2, 82 percent. Neither version gives a jailbreak score for Gemini Robotics 2.5 itself. Even the rules for measuring obedience, it seems, are still being drafted.
The full ASIMOV dataset is on Hugging Face for anyone building embodied AI. The model itself is not. Gemini Robotics 2.5 is not generally available. It has gone to Google’s robotics partners, among them Agile Robots, Apptronik and Boston Dynamics.