DeepMind has unveiled Gemini Robotics 2.5, demonstrating significant improvements in dexterous manipulation, including tying bows with deformable ribbons — a previously difficult task for robots. The system also achieved 30–100% higher success rates and 20–30% faster execution compared to its predecessor, with live deployments at Mercedes-Benz factories. Additionally, DeepMind launched the open safety benchmark ASIMOV-jailbreak to evaluate robotic AI safety.
Background
Google DeepMind's Gemini Robotics platform combines vision-language models with action generation to enable humanoid robots to perform complex physical tasks. This release follows ongoing industry efforts to bring general-purpose robot manipulation closer to human-level performance.
- Source
- Ars Technica
- Published
- Nov 10, 2026 at 11:00 PM
- Score
- 7.0 / 10