Scientific American reports DeepMind’s Gemini Robotics 2 and Asimov benchmark for robot safety
DeepMind is testing multimodal vision-language-action models (Gemini Robotics ER 2) on whole-body humanoids and introduced the Asimov benchmark to evaluate robot behavior and safety considerations.
In this brief: 3 sections 1 min read
ER 2 decomposes verbal requests into sequences of steps
Lower-level models convert visual input and plans into continuous movement
Works across different robot platforms including Apptronik’s Apollo 2
Safety measures include safe-stop behavior when humans approach
New Asimov benchmark created to assess robot behavior and potential for harm
Robot bodies also carry hardware safety systems (sensors, limits) alongside model-level protections
Demos include fetching objects, changing lightbulbs, tying knots
Researchers highlight limitations around tactile sensing and pressure estimation
Work described as early-stage but progressing toward whole-body control