DeepMind releases Gemini Robotics 2.5 and an open robot-jailbreak safety benchmark
Apptronik's Apollo 2 humanoids running the model kitted car parts at a 96% success rate in a Mercedes-Benz pilot, and the ASIMOV-Jailbreak set of 1,500 unsafe prompts is on Hugging Face.
Why it matters
DeepMind says the new vision-language-action model is 30 to 100 percent more successful and 20 to 30 percent faster than the previous version in lab tests, and that Gemini 4 Argon generated simulated practice runs that sped up training. The 96 percent factory result still trails a human, and DeepMind's director called closing the last 4 percent a steep road. There is no agreed way to test physical AI for safety; with Argon in the loop the system scored 98.8 percent on the jailbreak set, versus 73.5 percent for Gemini Robotics ER 2 alone. The model still runs on cloud servers and is not generally available.
Key facts
- 30% to 100% higher success rate than the previous version
- 20% to 30% faster in lab tests
- 96% success kitting car parts in a Mercedes-Benz pilot with Apptronik
- ASIMOV-Jailbreak holds 1,500 unsafe prompts
- 98.8% jailbreak safety with Argon in the loop, 73.5% for ER 2 alone
- Not generally available; runs on cloud servers