Abstract:
Google DeepMind released the vision-language-action (VLA) model Gemini Robotics 2.5 on November 10, local time, calling it the most powerful VLA model to date. The model is already available to Google's robotics partners.
According to the official blog, the team evaluated the model in more than 200 real tasks, including dividing meals with food tongs, tying bows on gift boxes, and pipetting liquids. Officials say that its mission success rate has increased by 30% to 100% compared to the previous baseline (depending on the mission type), and its completion speed has increased by 20% to 30%. The above is laboratory evaluation data.
In terms of implementation pilots, DeepMind cooperated with Apptronik and Mercedes-Benz to conduct a proof of concept on the in-use automobile production lines in Kecskemet, Hungary and Sindelfingen, Germany: Apptronik's Apollo 2 humanoid robot autonomously completed the task of sorting automobile parts (kitting) - picking metal brackets from bulk bins in production line order and placing them into corresponding sorting containers. According to officials, the success rate of this pilot task reached 96%, which was jointly achieved by Gemini Robotics ER and 2.5 dual models: ER is responsible for tracking the overall workflow, and 2.5 is responsible for controlling arm and hand movements. The pilot is currently in the proof-of-concept stage, and the official has not disclosed engineering indicators such as task duration, number of tests, and production line rhythm.
In terms of security, DeepMind simultaneously released the open source test benchmark ASIMOV-Jailbreak, which contains hundreds of prompt words disguised as normal requests and covers dozens of scenarios. It is used to test jailbreak attacks against robots (that is, bypassing security constraints through adversarial instructions). According to the official blog, Gemini 4 Argon's anti-jailbreak rate in this set of tests reached 99.5%, which is significantly higher than the 82% of the previous generation baseline Gemini Robotics ER 2. This set of results belongs to Argon, not Robotics 2.5 released this time.
In terms of openness, the model is currently only open to Google's robot partners. Whether and when it will be open to the public has not been officially stated.
Comments