Gemini Robotics 2.5 Debuts with Factory Trials and Open ASIMOV-Jailbreak Benchmark
DeepMind's Gemini Robotics 2.5 doubles success rates on some dexterous tasks and reached a 96% success rate kitting parts in Mercedes-Benz factories, though closing the remaining gap will be a 'steep road.' The model still runs on cloud servers with latency, and training relied on Gemini 4 Argon-generated synthetic data combined with real demonstrations. DeepMind released the open ASIMOV-Jailbreak dataset of 1,500 unsafe prompts, on which Argon-in-the-loop scores 98.8%.
Skynet Chance (+0.04 pp): Broader physical capability for general robots increases the potential for real-world harm. The open jailbreak safety benchmark and hazard-avoidance reasoning provide meaningful mitigations.
Skynet Date (1 day sooner): Deployment to partners such as Boston Dynamics and use in factories accelerates the move of AI into physical environments, though latency and reliability gaps temper the pace.
AGI Progress (+0.04 pp): Generalist manipulation of deformable objects and real-world deployment show substantial progress in physical intelligence, while noted limitations in latency and reliability keep it short of human level.
AGI Date (1 day sooner): Simulation-based data generation by frontier models eases the core robotics data bottleneck, speeding future progress.
<< All AI news for November 10, 2026
[ Get the daily index digest on Telegram → ]Related AI News
- Google DeepMind Launches Gemini Robotics 2.5 for General Real-World Dexterity
- Google Unveils Gemini 4 Argon Model for Robotics Applications
- Google DeepMind Launches Gemini 4 Argon Frontier Model with Leading Agentic, Scientific, and Embodied Reasoning Capabilities
- Ex-OpenAI Researcher Launches "Jev," a Non-Language Transformer That Outputs Calibrated Probabilities
- Startups Turn to Brain Waves and Muscle Sensors to Break the Robotics Data Bottleneck