Editorial illustration for DeepMind Reveals Gemini Robotics 2.5 and Open Safety Benchmark ASIMOV-Jailbreak
DeepMind's Gemini Robotics 2.5 Passes Factory Floor Test
DeepMind Reveals Gemini Robotics 2.5 and Open Safety Benchmark ASIMOV-Jailbreak
Google DeepMind put a new version of its robot brain to work on an actual Mercedes-Benz factory floor, not just in a lab, and the company is framing it as proof that general-purpose robotics is closer to matching human performance than most people assume. The update, Gemini Robotics 2.5, is the action-generating half of DeepMind's two-part system: a vision-language model called Gemini Embodied Reasoning handles planning from video and instructions, while Gemini Robotics turns that plan into actual motor commands.
DeepMind ran the new model through more than 200 tasks, from plating food with tongs to handling a lab pipette, and reported success rates 30 to 100 percent higher than the previous version, with execution speeds up 20 to 30 percent. Apptronik's Apollo 2 humanoid, fitted with the model, managed to tie a bow using its five-fingered hands, a task that trips up robots because ribbon and other soft, deformable material don't behave predictably under a gripper.
Alongside the model, DeepMind released an open safety benchmark, ASIMOV-jailbreak, aimed at the harder problem waiting behind all this dexterity: what happens when these systems start working next to people.
In the 2.5 release, Google says its robotic AI is getting tantalizingly close to completing real-world tasks as well as humans do. In fact, some Gemini robots were able to get actual work done in Mercedes-Benz factories, and the team is preparing for a day when robots operate alongside humans with an open source safety benchmark.
Why this matters
A Mercedes-Benz factory floor is a decent proving ground, but it's still one plant with controlled conditions, and the gap between "worked on the line in Stuttgart" and "safe to deploy broadly" is exactly what ASIMOV-jailbreak is meant to probe. We'd watch how quickly other labs and manufacturers adopt that benchmark rather than inventing their own; a fragmented safety-testing landscape helps nobody, least of all the people who'll eventually share a workspace with these machines.
The layered architecture, Gemini Embodied Reasoning handling the planning on top of live video, is worth tracking too. Stack enough models between a verbal instruction and a physical action, and you get more points where things can go wrong, not fewer. DeepMind open-sourcing a jailbreak benchmark suggests they know that.
For founders building on Gemini Robotics, the real test isn't the demo reel, it's whether this reasoning layer holds up when instructions are ambiguous or the room isn't a factory DeepMind helped design. That's the next thing to watch, not the headline numbers.
Related Reading
Browse all Google news →
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
Related Topic Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI
Related Topic NVIDIA and Google Cloud let developers scale AI from prototype to production
Market Trends
From Archives AI Agents Test Zero-Trust Limits as Single PEP Scaling Falls Short
From Archives Meta Releases Muse Ember, Lightweight Agent Model at 10 Cents Per Million Tokens
From Archives Gemini 4 Carbon arrives in fastest model cycle; Google says Argon better
LLMs & Generative AI
From Archives Google launches two TPUs with confidential computing, 2x lower latency
Business & Startups
Common Questions Answered
How does Gemini Robotics 2.5 differ from previous versions in real-world factory applications?
Gemini Robotics 2.5 has demonstrated the ability to complete actual work tasks on Mercedes-Benz factory floors, moving beyond laboratory testing to prove that general-purpose robotics can match human performance in real-world conditions. This represents a significant advancement in DeepMind's two-part system, where Gemini Embodied Reasoning handles planning from video and instructions while Gemini Robotics executes those plans into actual robotic actions.
What is the ASIMOV-Jailbreak benchmark and why is it important for robot deployment?
ASIMOV-Jailbreak is an open-source safety benchmark developed by DeepMind to test robotic AI systems before they operate alongside humans in shared workspaces. The benchmark is designed to bridge the gap between successful factory testing and safe broad deployment, ensuring that safety testing remains standardized across labs and manufacturers rather than fragmented by individual proprietary benchmarks.
What are the two main components of DeepMind's robotic AI system?
DeepMind's system consists of Gemini Embodied Reasoning, a vision-language model that handles planning from video and instructions, and Gemini Robotics, which converts those plans into actual robotic actions. Together, these two components enable the robot brain to both understand tasks and execute them in real-world environments like factory floors.
Why is the Mercedes-Benz factory floor test considered significant but not sufficient for broader deployment?
While the Mercedes-Benz factory floor provides a legitimate proving ground for Gemini Robotics 2.5, it represents a single plant with controlled conditions rather than the diverse, unpredictable environments robots would encounter in broader deployment. The gap between successful performance in one controlled factory setting and safe operation across multiple facilities and workspaces is precisely what the ASIMOV-Jailbreak safety benchmark is designed to evaluate.
Further Reading
- Papers with Code Benchmarks - Papers with Code
- Chatbot Arena Leaderboard - LMSYS
