DeepMind unveils Gemini Robotics 2.5 and ASIMOV-jailbreak benchmark
Original title: DeepMind reveals Gemini Robotics 2.5 and open safety benchmark "ASIMOV-jailbreak"
Apollo 2 robots kitted parts in Mercedes factories at 96% success; the open jailbreak set has 1,500 prompts.
Google DeepMind released Gemini Robotics 2.5, a vision-language-action model it says is 30 to 100 percent more successful and 20 to 30 percent faster than the prior version on more than 200 lab tasks. With Apptronik, Gemini-powered Apollo 2 robots kitted automotive parts in a few European Mercedes-Benz factories at a 96 percent success rate. DeepMind also opened ASIMOV-Jailbreak, 1,500 prompts meant to elicit unsafe robot actions, reporting 98.8 percent safety with Gemini 4 Argon in the loop versus 73.5 percent with Gemini Robotics ER 2 alone. The 2.5 release still runs on cloud servers.
Unconfirmed: Success-rate and jailbreak-safety figures are Google’s own lab and factory measurements, not an independent evaluation.
- Apptronik’s Apollo 2 with Gemini Robotics 2.5 was shown tying a bow.
- Obstacle handling raised safe task completion 33 percent, and 43 percent with people nearby.
- Gemini 4 Argon generated simulation variants from real camera data for training.
- ASIMOV-Jailbreak is published on Hugging Face; general availability of 2.5 is not stated.
- Vendors
- Products
- Apollo 2
- Organizations
- Google DeepMindApptronikMercedes-Benz
From the original article
5 min read at arstechnica.com
Two Gemini-powered robots hard at work in a Mercedes-Benz factory. Two Gemini-powered robots hard at work in a Mercedes-Benz factory. Google DeepMind has revealed a new generation of the AI-powered Gemini Robotics platform.
By Ryan Whitwam · arstechnica.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.