Static

DeepMind reveals Gemini Robotics 2.5 and open safety benchmark "ASIMOV-jailbreak"

First reported by Ars Technica ·

The signal ●○○○ Compiled by AI from Ars Technica, the single source so far
Why you might care

Robots can now perform complex manipulation tasks with increased success and speed, moving closer to human-level dexterity.

What happened

Google DeepMind has released Gemini Robotics 2.5, an advanced AI platform for robots that demonstrates a significant improvement in dexterity and speed. The new model shows a 30-100% higher success rate and 20-30% faster execution on various tasks, including complex manipulation like tying a bow with a humanoid robot. In real-world trials at Mercedes-Benz factories, Gemini-powered robots achieved a 96% success rate when kitting automotive parts. DeepMind also introduced ASIMOV-Jailbreak, an open-source safety benchmark with 1,500 prompts designed to test and improve the safety of robotic AI by identifying potentially hazardous behaviors. The company is collaborating with robotics partners like Apptronik and Boston Dynamics to deploy this technology, aiming for robots that can operate safely alongside humans.

What it means

The advancements in Gemini Robotics 2.5, particularly its success in handling deformable objects and achieving high accuracy in factory settings, signal a maturing of AI in physical automation. This progress, accelerated by generative AI simulation tools like Gemini 4 Argon, indicates a shift towards more generalized robotic capabilities rather than task-specific programming. While still facing latency due to cloud-based processing, the improvements in safety features, such as inference-time reasoning to avoid obstacles and humans, are crucial for real-world deployment.

The introduction of the ASIMOV-Jailbreak benchmark by DeepMind is a critical step towards establishing industry-wide safety standards for embodied AI. By providing an open dataset of adversarial prompts, DeepMind aims to foster more robust and reliable robotic systems, mitigating risks associated with AI failures. This move encourages greater transparency and collaboration in the field, pushing the entire robotics sector toward more secure and predictable AI integration, especially in human-robot collaborative environments.

AI-written summary. May contain errors.

Google Robotics