DeepMind Releases Gemini Robotics 2.5 with 96% Factory Success Rate

Google DeepMind released Gemini Robotics 2.5, a vision-language-action model for physical AI, on November 9-10, 2026. The model achieves 30-100% higher task success rates and completes tasks 20-30% faster than previous baselines across more than 200 distinct tasks. In a proof-of-concept deployment, Apptronik's Apollo 2 humanoid robots running the model achieved a 96% success rate on automotive kitting tasks at Mercedes-Benz production lines in Hungary and Germany. DeepMind described the result as "promising progress toward widespread real-world industrial applications."

The platform separates reasoning, handled by Gemini Embodied Reasoning, from action generation, handled by Gemini Robotics, in distinct computational layers. Demonstrated capabilities include folding laundry, serving pastries, sorting car parts on a factory floor, and tying a bow with five-fingered hands. DeepMind noted that deformable objects like ribbon are notoriously difficult for robots to manipulate.

New inference-time safety reasoning boosted safe task completion around obstacles by 33% and around nearby people by 43%. Gemini 4 Argon served as a simulation engine, generating thousands of varied simulated practice runs from real-world camera scenes to build digital twins. On the ASIMOV-Jailbreak benchmark, Gemini 4 Argon achieved 99.5% jailbreak resistance compared to 82% for the prior baseline, Gemini Robotics ER 2. DeepMind released ASIMOV-Jailbreak as an open-source benchmark intended to support safe human-robot coexistence.

Separately, Google had made Gemini 2.5 Flash and Pro embodied reasoning capabilities available for robotics applications via Google AI Studio, the Gemini API, and Vertex AI as of June 2025, with a trusted tester program for a specialized Gemini Robotics-ER variant.