Not established from the available sources.
GOOGLE DEEPMIND · MODEL RELEASE TRACKER
Gemini Robotics 2.5
Gemini Robotics 2.5 is a vision-language-action (VLA) robot control model from Google DeepMind, announced 2026-11-10. It expands task breadth, dexterity, reliability, and safety for real-world robotic manipulation and is being rolled out to robotics partners.CURRENT SNAPSHOT3/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Not established from the available sources.
Model modalityDescribed as a vision-language-action (VLA) model that processes visual information, understands language instructions, and translates them into robot movements.
Production pilot success rateIn a pilot on active Mercedes‑Benz production lines (Apptronik Apollo 2 humanoid), reported a 96% success rate for sorting metal parts and kitting tasks.
Availability to partnersGemini Robotics 2.5 is rolling out to DeepMind's robotics partners.
RELEASE TIMELINE
Published, source-backed release events only.
VERIFIABLE FACTS
Every value stays attached to a source and date.
RELEASE · Announcement and rolloutDEVELOPER CLAIM
Announced by Google DeepMind on 2026-11-10 and rolling out to robotics partners.
MODALITIES · Model modalityDEVELOPER CLAIM
Described as a vision-language-action (VLA) model that processes visual information, understands language instructions, and translates them into robot movements.
CAPABILITIES · CapabilitiesDEVELOPER CLAIM
Increases the breadth of skills robots can learn and the speed at which they acquire them; expands task breadth, performance, and complexity; improves dexterity, reliability, and safety relative to prior robotics models.
CAPABILITIES · Evaluation scopeDEVELOPER CLAIM
Evaluated across over 200 distinct tasks (examples cited: serving food with tongs, tying a bow on a gift box, pipetting a liquid) with reported significant increases in task precision and dexterity.
CAPABILITIES · Real-world proof-of-conceptDEVELOPER CLAIM
In a real-world proof-of-concept on Mercedes‑Benz production lines, models using Gemini Robotics 2.5 helped Apptronik’s Apollo 2 humanoid pick and pack automotive parts with a 96% success rate.
AVAILABILITY · Availability to partnersDEVELOPER CLAIM
Gemini Robotics 2.5 is rolling out to DeepMind's robotics partners.
SAFETY · Safety steerability and adversarial jailbreak benchmarkDEVELOPER CLAIM
Reports improved physical safety steerability and the introduction/release of a benchmark (ASIMOV-Jailbreak) to test adversarial jailbreaks specific to robotics.
CAPABILITIES · Task breadth and typesDEVELOPER CLAIM
Evaluated across over 200 distinct tasks (examples: serving food with tongs, tying a bow, pipetting liquids), expanding task breadth, performance, and complexity to improve dexterity and handling of deformable and delicate objects.
CAPABILITIES · Relative performance vs previous baselinesDEVELOPER CLAIM
Reported to perform 30–100% better than DeepMind's previous baselines depending on task type, and to complete tasks 20–30% faster.
BENCHMARKS · Production pilot success rateDEVELOPER CLAIM
In a pilot on active Mercedes‑Benz production lines (Apptronik Apollo 2 humanoid), reported a 96% success rate for sorting metal parts and kitting tasks.
WHAT CHANGED
Stored passport versions, without reconstructed history.
RELEASE · Announcement− Gemini Robotics 2.5 was introduced/announced on 2026-11-10.
RELEASE · Announcement and rollout+ Announced by Google DeepMind on 2026-11-10 and rolling out to robotics partners.
AVAILABILITY · Availability− Gemini Robotics 2.5 is rolling out to DeepMind’s robotics partners.
MODALITIES · Modality− Described as a vision-language-action (VLA) model for robotics.
AVAILABILITY · Availability to partners+ Gemini Robotics 2.5 is rolling out to DeepMind's robotics partners.
BENCHMARKS · Production pilot success rate+ In a pilot on active Mercedes‑Benz production lines (Apptronik Apollo 2 humanoid), reported a 96% success rate for sorting metal parts and kitting tasks.
CAPABILITIES · Relative performance vs previous baselines+ Reported to perform 30–100% better than DeepMind's previous baselines depending on task type, and to complete tasks 20–30% faster.
SAFETY · Safety steerability and adversarial jailbreak benchmark+ Reports improved physical safety steerability and the introduction/release of a benchmark (ASIMOV-Jailbreak) to test adversarial jailbreaks specific to robotics.
SAFETY · Safety and benchmark− DeepMind reports improved physical safety steerability and the introduction of a new benchmark to test against adversarial jailbreaks specific to robotics.
CAPABILITIES · Task breadth and types+ Evaluated across over 200 distinct tasks (examples: serving food with tongs, tying a bow, pipetting liquids), expanding task breadth, performance, and complexity to improve dexterity and handling of deformable and delicate objects.
MODALITIES · Model modality+ Described as a vision-language-action (VLA) model that processes visual information, understands language instructions, and translates them into robot movements.