Google DeepMind Releases Gemini Robotics 2.5, Claims 96% Success on Mercedes-Benz Factory Pilot

Google DeepMind announced Gemini Robotics 2.5 on November 10, calling it its most capable vision-language-action (VLA) model to date. The model is rolling out to robotics partners — not the public — after a proof-of-concept run on Mercedes-Benz production lines where Apptronik's Apollo 2 humanoid robots picked and packed automotive parts with a reported 96% success rate.

What shipped

Gemini Robotics 2.5 is a VLA model: it takes visual input and language instructions and outputs robot actions. DeepMind says it improves on three axes — dexterity, reliability, and safety — and that Gemini 4 Argon, a separate Gemini model, accelerated its development by generating simulated practice runs from real camera scenes.

The announcement names three partners: Agile, Apptronik, and Mercedes-Benz. No price, public API, or technical paper was released. DeepMind's blog post states the model "is rolling out to our robotics partners."

The Mercedes-Benz pilot

The headline reliability figure comes from a proof of concept at Mercedes-Benz plants in Kecskemét, Hungary, and Sindelfingen, Germany. Apollo 2 robots performed automotive kitting: reaching into bulk bins, grasping metal brackets, and placing them into sequencing trays in vehicle-build order.

Two models divided the work. Gemini Robotics ER (embodied reasoning) tracked the multi-step workflow. Gemini Robotics 2.5 controlled the robot's hands and arms. DeepMind reports a 96% success rate on this task.

DeepMind director Kanishka Rao told Ars Technica that closing the remaining 4% gap would be a "steep road." Robotics head Carolina Parada described a tension between making a robot excellent at one application and preserving its ability to handle many different jobs, saying that pushing for generalization may mean not reaching 99.9% reliability in every application [6].

The post does not say how many picks the 96% covers, nor how long the pilot ran. It is labeled a proof of concept, not a production deployment.

Lab results: 200 tasks, unnamed baselines

DeepMind evaluated the model on "over 200 distinct tasks" including serving food with tongs, tying a bow on a gift box, and pipetting liquid. It reports a 30% to 100% higher success rate than "previous baselines" depending on task type, and 20% to 30% faster completion.

The baselines are not named, and their absolute scores are not given. The post says the model works on Apptronik's Apollo 2 (with a 22-degree-of-freedom SharpaWave hand) and the Franka Duo dual-arm platform with two-finger grippers.

Ars Technica reports the model still runs on cloud servers, meaning the reported speed gains do not establish low-latency, on-robot operation.

Safety: steerability and a new jailbreak benchmark

DeepMind says Gemini Robotics 2.5 uses inference-time reasoning guidance to steer around hazards in real time. In its evaluations, this boosted safe task completion by 33% around obstacles and 43% around nearby people.

The company is also releasing ASIMOV-Jailbreak, an open-source benchmark of hundreds of prompts that disguise unsafe requests across dozens of scenes. DeepMind reports 99.5% jailbreak resistance for Gemini 4 Argon on this benchmark, compared to 82% for Gemini Robotics ER 2.

The 99.5% figure belongs to Argon, not to Gemini Robotics 2.5 itself. The post gives no ASIMOV-Jailbreak score for the VLA model being released.

What is not in the release

• No public access, pricing, or API documentation.
• No named baselines for the 30–100% improvement claim.
• No sample size for the 96% factory figure.
• No technical paper for Gemini Robotics 2.5 (a Gemini 4 safety report is linked).
• Every number comes from DeepMind or its partners; no independent evaluation is cited.

What to watch

• Whether ASIMOV-Jailbreak scores from other labs match DeepMind's.
• Whether the Mercedes pilot extends to sustained, shift-long operation with recovery from grasp failures.
• Whether partners beyond the named three receive access, and on what terms.
• Whether on-robot compute replaces cloud inference for latency-sensitive tasks.

DeepMind calls Gemini Robotics 2.5 "an important milestone on the path toward solving physical AGI." The factory pilot is a concrete step. The gap between a 96% proof of concept and a robot that stays on the line remains, as DeepMind's own researchers acknowledge, a steep road.