On November 10, 2026, Google DeepMind officially released its most capable Vision‑Language‑Action (VLA) model to date—Gemini Robotics 2.5. The model is designed to translate visual perception and natural‑language instructions directly into robot‑executable motion commands, further improving robots' general dexterity and reliability in the real physical world. According to DeepMind, Gemini Robotics 2.5 shows significant advances in task breadth, learning speed, fine manipulation and reliability, and has completed proof‑of‑concept tests in actual factory environments.
Core capability upgrades and evaluation performance
Gemini Robotics 2.5 is positioned as "the most advanced VLA model to date," capable of controlling a range of platforms from dual‑arm robots to full humanoids. It works in tandem with the Gemini Robotics ER (Embodied Reasoning) model, which handles high‑level planning: the ER model interprets instructions and live video to devise strategies, while the VLA model converts plans into concrete joint velocities, poses and gripper control commands.
The team evaluated the model on over 200 different tasks, covering fine‑manipulation scenarios such as using tongs to distribute food, tying bows on gift boxes, and handling liquids with pipettes. These tasks collectively test tool use, manipulation of deformable objects and handling of fragile items. Results show that, compared with prior baselines, Gemini Robotics 2.5 improves task success rates by 30% to 100% (depending on task) and speeds up completion by 20% to 30%. Some previously nearly impossible tasks showed dramatic gains—for example, success in packaging pastries with tongs rose from about 5% to 100%, closing a toolbox lid went from 0% to 93%, and untying a ribbon bow rose from 0% to 70%.
In real industrial settings, Gemini Robotics 2.5 was demonstrated in collaboration with Apptronik’s Apollo 2 humanoid on Mercedes‑Benz production lines in Kecskemét, Hungary and Sindelfingen, Germany. The robot autonomously performed parts picking and kitting tasks—taking metal parts from bulk bins and placing them into trays in the order required for vehicle assembly—with a 96% success rate. DeepMind notes this can help shift repetitive work to robots so technicians can focus on more complex assembly tasks.
The model supports multiple robot morphologies, including the Apollo 2 humanoid equipped with a five‑finger, 22‑DOF SharpaWave hand, and the Franka Duo platform using a two‑finger parallel gripper. It provides whole‑body intelligent control and can dynamically adjust the center of mass so robots can safely step, crouch or bend to operate in complex environments.
Training acceleration and safety considerations
DeepMind says Gemini 4 Argon played a key role in training Gemini Robotics 2.5. Argon can generate virtual practice environments from single‑view camera footage and replay failure cases to refine strategies, significantly shortening the physical data‑collection cycle and enabling the model's capability jump in roughly four months.
On safety, the team improved obstacle avoidance and human–robot coexistence behaviors. Official data indicate about a 33% improvement in safe task completion near obstacles and about a 43% improvement near people. DeepMind also released the open‑source safety benchmark ASIMOV‑Jailbreak to evaluate models' resistance to "dangerous requests disguised as harmless instructions." The associated dataset has been made available to the research community.
Objective assessment and industry impact
The release of Gemini Robotics 2.5 marks a substantive step for VLA models from lab demonstrations toward real industrial deployment. A 96% factory kitting success rate and multi‑task improvements suggest the model already has strong practical utility in structured, repeatable industrial environments. Advances in fine manipulation—especially handling of deformable and fragile objects—also narrow the gap between robots and humans on everyday dexterous tasks.
However, an objective view is warranted. Current evaluation data come mainly from official and partner tests; large‑scale, long‑term independent validation remains limited. Although the factory proof‑of‑concept reached 96% success, failure modes in mass production, long‑duration operational stability and close collaboration details with human workers still require further observation. The model is currently made available primarily to DeepMind's robotics partners; no public API, pricing or broad commercial rollout has been announced.
From an industry perspective, Gemini Robotics 2.5 reinforces the "general physical intelligence" approach: a unified VLA architecture adaptable to different robot bodies, combined with large‑model‑accelerated simulation and learning. If high reliability can be maintained across more scenarios and issues around safety, cost and deployment barriers are addressed, this could accelerate the move of humanoid and dual‑arm robots from demonstration to practical production assistance. The release of an open safety benchmark also provides a reference tool for assessing embodied AI risks in the industry.
Overall, Gemini Robotics 2.5 represents an important advance in balancing dexterity and real‑world applicability for current VLA technology. It demonstrates rapid iteration of model capabilities while underscoring that engineering and safety challenges remain between laboratory success rates and long‑term factory operation.
▶️ Enjoy this article? Creating content is hard!
Follow our YouTube channel:
AI Rosetta
Your likes, subscriptions and shares are our greatest support! 🙏
For more AI news visit:
https://www.airosetta.com