Google DeepMind's Gemini Robotics 2.5 Hits 96% at Mercedes-Benz Factory Lines
DeepMind's two-model design pairs motion control with a reasoning layer, but cloud dependency caps scale

Google DeepMind's Gemini Robotics 2.5 has completed live production trials at Mercedes-Benz factories in Hungary and Germany, achieving a 96% success rate on automotive kitting tasks when deployed through Apptronik's Apollo 2 humanoid robot. The result, confirmed at the Conference on Robot Learning (CoRL) in Austin on November 10, represents the most commercially significant field evaluation DeepMind has published for its vision-language-action model family — but the same deployment revealed a structural limitation that will define the technology's near-term ceiling: GR 2.5 still runs entirely on cloud servers, constraining the throughput and reliability that full-scale factory operations demand.
A Two-Model Architecture Built for Manufacturing Complexity
Gemini Robotics 2.5 is not a single system. DeepMind's current deployment architecture pairs two specialized models: GR 2.5 itself handles arm and hand motion control, translating high-level instructions into precise joint-level actuation commands; Gemini Robotics ER (Embodied Reasoning) manages task-level planning, tracks whether individual steps have succeeded, and replans adaptively when a manipulation attempt fails mid-execution.
This division of labor reflects a genuine engineering constraint. Robotic actuators require motor commands at rates in the hundreds of hertz, demanding fast, deterministic inference. Language-grounded task planning, by contrast, benefits from broader contextual reasoning at slower timescales. By separating the two functions into specialized models, DeepMind can optimize each independently rather than forcing a single model to satisfy both real-time motor control and deliberate planning simultaneously.
A third component, Gemini 4 Argon, sits behind the system as a simulation and safety layer. Argon analyzes real camera footage from robot work cells, generates thousands of synthetic training variants of each scene — building a digital twin without requiring equivalent physical trials — and runs failure-mode analysis to improve GR 2.5's training pipeline. It is also the component responsible for Gemini Robotics' recently published safety evaluation: on the new ASIMOV-Jailbreak benchmark, an open-source test that uses hundreds of prompts designed to disguise unsafe instructions, Argon achieved 99.5% resistance. That figure describes Argon's safety behavior, not GR 2.5's manipulation capability — a distinction several secondary sources have elided.
Automotive Kitting as a Commercial Benchmark
The Mercedes-Benz trial focused on automotive kitting — sorting individual components from bulk supply bins into vehicle-assembly sequencing trays in the precise order the production line requires. The task combines several manipulation challenges simultaneously: visually recognizing thousands of distinct part types under varied lighting conditions, grasping objects of different shapes and weights with consistent force control, and placing components accurately within tight spatial tolerances.
Apptronik's Apollo 2 humanoid, which runs GR 2.5 in the Mercedes-Benz deployments, brings a five-finger hand designed to replicate the dexterous range that automotive assembly work requires. The trials ran across two facilities: Mercedes-Benz's Kecskemét plant in Hungary and its Sindelfingen complex in Germany, both active production sites rather than controlled test environments.
Across the 200-plus physical tasks DeepMind evaluated with GR 2.5, the company reports 30 to 100 percent higher success rates compared with the previous generation. All performance figures are company-reported; no independent benchmark reproduction has been published as of November 16, 2026. The 96% kitting success rate, while meaningful, also implies roughly one in twenty automotive components handled incorrectly — a failure rate that full production lines could tolerate only if downstream recovery systems exist.
Cloud Dependency Defines the Production Timeline
The fundamental constraint Gemini Robotics 2.5 faces is not its manipulation capability — it's where the inference happens. Kanishka Rao, a senior researcher on the GR 2.5 team, confirmed to Ars Technica that the system runs on DeepMind's cloud infrastructure rather than embedded processors on the robot itself. Network latency adds measurable delay to every inference cycle. At the controlled throughput of a pilot study, that overhead is manageable. At the volume of a full automotive production line — where thousands of component movements occur per hour and network interruptions carry immediate line-stop risk — it represents a reliability and throughput ceiling that the system has not yet overcome.
Apptronik and DeepMind have framed broad commercial deployment as "expected in 2027 and beyond," a timeline consistent with the remaining work required: edge inference capability, production-grade reliability validation, and regulatory compliance under the EU AI Act's high-risk workplace provisions, which apply to AI-driven robotic systems operating alongside humans in European factories.
Where GR 2.5 Sits Against Competing VLA Platforms
The VLA model landscape in late 2026 has bifurcated along a dimension that matters more for manufacturing customers than benchmark rankings: whether inference runs in the cloud or at the edge.
NVIDIA's GR00T N1.7, released earlier this year, runs on the company's Jetson Thor accelerator with on-device inference — significantly faster than any cloud-round-trip architecture can achieve. GR00T N1.7 carries an open commercial license and targets a broad OEM audience, trading some of GR 2.5's breadth and task variety for the deployment independence that factory operators require. Physical Intelligence's π0 maintains a closed, partner-access model with strong results in dexterous manipulation tasks, but similarly relies on proprietary cloud infrastructure. Figure AI's Helix 02 runs fully onboard its own humanoid platform, eliminating network dependency entirely at the cost of hardware lock-in.
GR 2.5's competitive position is clearest on task breadth and in environments where network connectivity can be guaranteed — controlled factory cells with dedicated bandwidth and existing Mercedes-Benz IT infrastructure, for instance. Its competitive exposure is sharpest wherever an operator requires guaranteed sub-100ms cycle times or cannot accept network-dependent reliability.
Inference-Time Safety and the Road to Wider Deployment
One mechanism built into GR 2.5 specifically addresses factory safety: inference-time reasoning guidance, which steers robot motion trajectories around classified physical hazards during live execution rather than relying entirely on pre-training. DeepMind reports a 33% improvement in safe task completion around fixed obstacles and a 43% improvement in safe behavior around people. The mechanism does not stop tasks — it reroutes them, allowing GR 2.5 to continue operating while navigating proximity to workers or equipment. This architecture matters for regulatory compliance: European workplace safety standards require demonstrable hazard-response capability, not merely pre-deployment testing.
The confirmed partner rollout for GR 2.5 extends beyond Apptronik and Mercedes-Benz. Boston Dynamics and Agile Robots are listed as additional deployment partners, suggesting DeepMind is intentionally building a diverse hardware base rather than concentrating on a single platform.
What comes next for Gemini Robotics is likely less about capability improvements than about the deployment architecture shift that would let those capabilities convert into commercial revenue. The Mercedes-Benz result establishes that GR 2.5 can meet automotive-grade task thresholds when conditions are controlled. The question the industry will watch through 2027 is whether DeepMind can deliver the same performance without the cloud dependency that currently makes scaling those conditions to full production still a commitment yet to be honored.



