RELEASE · MODELS · #3346
Google introduces Gemini Robotics 2.5, a more capable vision-language-action model for real-world dexterity
Google announced Gemini Robotics 2.5, an upgraded vision-language-action (VLA) model that increases task breadth, dexterity, reliability, and safety for robots. Google says the model was evaluated on over 200 tasks (showing 30–100% higher success rates and 20–30% faster completion versus prior baselines) and achieved a 96% success rate in a Mercedes‑Benz proof‑of‑concept using Apptronik’s Apollo 2 humanoid on production lines.
KEY POINTS
- Google announced Gemini Robotics 2.5, an upgraded vision-language-action (VLA) model that increases task breadth, dexterity, reliability, and safety for robots.
- Google says the model was evaluated on over 200 tasks (showing 30–100% higher success rates and 20–30% faster completion versus prior baselines) and achieved a 96% success rate in a Mercedes‑Benz proof‑of‑concept using Apptronik’s Apollo 2 humanoid on production lines.
- This matters because the release claims substantial gains in dexterity, speed, and safety and includes an industrial proof‑of‑concept that points toward wider real‑world deployment of robotic AI.
DECISION BRIEF
CONFIDENCE
WHAT CHANGED
Google DeepMind announced Gemini Robotics 2.5, an upgraded vision‑language‑action (VLA) model. According to the company, the model expands task breadth and improves dexterity, reliability, and safety; it was evaluated on over 200 distinct tasks with reported 30–100% higher success rates and 20–30% faster completion versus prior baselines. In a Mercedes‑Benz proof‑of‑concept using Apptronik’s Apollo 2 humanoid on active production lines, Google reports a 96% success rate. Google also says it improved physical safety steerability and introduced a benchmark for adversarial jailbreaks (ASIMOV‑Jailbreak), and that Gemini Robotics 2.5 is rolling out to robotics partners.
WHY NOW
The company claims substantial performance, speed, and safety gains and demonstrates those claims with an industrial proof‑of‑concept on active production lines. Those elements together—performance metrics across 200+ tasks plus a factory pilot—are the basis for assessing near‑term readiness for more real‑world use and partner rollouts.
WHO IS AFFECTED
Directly named in the sources: Google DeepMind and its robotics partners; Apptronik (Apollo 2 humanoid); Mercedes‑Benz production lines (Kecskemét and Sindelfingen cited in the media excerpt); users of VLA robot control on platforms such as Apollo 2 and Franka Duo.
CONFIRMED
- Google DeepMind announced Gemini Robotics 2.5. - The model was evaluated across over 200 distinct tasks. - Google reports 30–100% higher success rates and 20–30% faster task completion versus its previous baselines, varying by task type. - In a Mercedes‑Benz proof‑of‑concept on active production lines with Apptronik’s Apollo 2 humanoid, Google reports a 96% success rate for a kitting task. - Example tasks cited include serving food with tongs, tying a bow, and pipetting liquids. - DeepMind uses Gemini 4 Argon as a simulation engine for training, per the media excerpt. - Google reports improvements in physical safety steerability and introduced a benchmark to test adversarial jailbreaks for robotics. - The model is described as working with different robot types including Apptronik’s Apollo 2 and the Franka Duo platform. - Google says Gemini Robotics 2.5 is rolling out to robotics partners.
UNCERTAIN
Missing or unclear in the supplied excerpts: - Independent verification: no third‑party or peer‑reviewed validation of the reported success rates, speed gains, or safety metrics is provided in the excerpts. - Evaluation details: the excerpts do not provide full methodology (exact task definitions, trial counts, success criteria, statistical uncertainty, or dataset composition). - Long‑term and scale stability: whether the reported performance and 96% pilot success persist over extended continuous production runs or across different factories is not specified. - Generalizability: extent of performance across robot hardware, part types, and unobserved edge cases beyond the cited examples (Apollo 2, Franka Duo) is unspecified. - Accessibility and rollout: precise scope, timing, licensing, or commercial availability to partners beyond the statement that it is rolling out is not provided. - ASIMOV‑Jailbreak details and uptake: the content, evaluation results, and external adoption of the benchmark are not detailed in the excerpts. No conflicting evidence appears in the provided excerpts; uncertainties are due to missing details rather than explicit contradictions.
WHAT TO WATCH
Concrete observable signals to monitor next (based on supplied excerpts): - Publication of full evaluation details (methodology, per‑task trial counts, success criteria) or peer‑reviewed papers validating the 200+ task results. - Independent third‑party tests or audits reporting performance on the same or similar tasks. - Announcements of additional production deployments or broader rollouts by Mercedes‑Benz, Apptronik, other manufacturers, or Google’s robotics partners. - Release, repository, or benchmark results for ASIMOV‑Jailbreak and uptake by external researchers. - Technical disclosures about integration with Gemini 4 Argon or open‑source artifacts used for training/simulation. - Performance logs or reports showing long‑run uptime, failure modes, or safety incident statistics in production settings.
WHY IT MATTERS
This matters because the release claims substantial gains in dexterity, speed, and safety and includes an industrial proof‑of‑concept that points toward wider real‑world deployment of robotic AI.
EVIDENCE MAP
4Editorial claims linked to specific sources, with support, contradiction and context shown separately.
Google DeepMind introduced Gemini Robotics 2.5, a new, more capable vision-language-action (VLA) model.
SUPPORTEDChecked 2026-11-10 · 2 supporting
The model was evaluated on over 200 distinct real‑world tasks (examples cited include serving food with tongs, tying a bow, and pipetting).
SUPPORTEDChecked 2026-11-10 · 2 supporting
DeepMind reports 30–100% higher success rates and 20–30% faster completion versus its prior baselines for different task types.
SUPPORTEDChecked 2026-11-10 · 1 supporting
The media excerpt explicitly reports the 30–100% higher success rates and 20–30% faster completion figures versus prior baselines.
CONTEXT Google DeepMindMEDIA · 2026-11-10DeepMind’s blog describes significant increases in task performance and dexterity but the excerpted text does not display the numeric ranges shown in the media report.
In a Mercedes‑Benz proof‑of‑concept on active production lines, Apptronik’s Apollo 2 humanoid achieved a reported 96% success rate sorting/picking and packing metal parts.
SUPPORTEDChecked 2026-11-10 · 2 supporting
DeepMind’s blog reports a 96% success rate for Apptronik’s Apollo 2 picking and packing automotive parts in a real‑world proof‑of‑concept on Mercedes‑Benz lines.
SUPPORTS The DecoderMEDIA · 2026-11-10The media report gives details of the pilot (locations, task description, two‑model setup) and cites the reported 96% success rate.
SOURCES & TIMELINE
2We’re enabling robots to learn faster for the real world - from everyday skills to the factory floor Today, we are introducing Gemini Robotics 2.5, our most capable vision-language-action (VLA) model to date. Gemini Robotics 2.5 increases both the breadth of skills a robot can learn and the speed at which it can acquire them. This enhances dexterity, reliability, and safety to help robots step out of the lab and into…
Google Deepmind introduced Gemini Robotics 2.5, a new VLA model it says performs 30 to 100 percent better than its own previous baselines depending on the task type. On active Mercedes-Benz production lines in Hungary and Germany, Apptronik's humanoid robot Apollo 2 sorted metal parts with a reported success rate of 96 percent. Google Deepmind uses its language model Gemini 4 Argon as a simulation engine for traini…