Overview
- Gemini Robotics 2.5, which DeepMind unveiled on Tuesday, November 10, 2026, is a vision‑language‑action model that the company reports achieves 30–100% higher success rates and 20–30% faster completion times across more than 200 tested tasks.
- DeepMind ran a factory proof‑of‑concept with Apptronik’s Apollo 2 humanoid on Mercedes‑Benz lines in Kecskemét and Sindelfingen that reached a 96% success rate on autonomous pick‑and‑pack automotive kitting tasks.
- The company released ASIMOV‑Jailbreak, an open safety benchmark of adversarial prompts, and reported that Gemini 4 Argon scored 99.5% jailbreak resistance compared with an 82% baseline on prior models.
- The system separates planning and motion by using Gemini Embodied Reasoning as the planner and the VLA model as the mover, and DeepMind says Gemini 4 Argon accelerated training by creating digital‑twin simulations from real camera scenes.
- DeepMind is rolling the software to partners such as Apptronik and Agile Robots to let hardware makers focus on mechanics, though the performance and safety gains are company‑reported and have not yet been widely independently verified.