Google DeepMind has introduced a robotics AI called Gemini Robotics-ER 1.5, which, for the first time, combines embodied reasoning with perception to let robots make multi-step decisions before acting, tells ArsTechnica. This is more than a control model; it’s meant to think about context, use memory, and plan. Accompanying it is Gemini Robotics 1.5, focused on vision, language, and action, and together they form a system that bridges thinking and doing.
At the core is a vision-language-action architecture. The “ER” model reasons about goals and context, while the 1.5 version handles perception and translating instructions into actuation. DeepMind says they fine-tuned these models using robotics data and benchmarks to give them stronger generalization across tasks and environments. They evaluated performance on more than a dozen embodied reasoning benchmarks, such as pointing, spatial reasoning, and task progress estimation, and claim state-of-the-art results.
One notable advance is motion transfer: behaviors learned for one robot can transfer to others, reducing the need to retrain for every specific robot architecture. Future robots can tap into this cross-platform learning. Also, the new models can use web search to fetch or supplement information when planning tasks.
That said, DeepMind doesn’t suggest these robots are sentient. The models still depend on training data, execution limitations, and constraints of hardware. Challenges remain in handling uncertainty, fragile real-world feedback, and safety.
This announcement marks a turning point: robots that do more than follow instructions can anticipate, reason, and adapt. It edges us closer to machines that operate in human environments more robustly. Whether those robots ever truly “think” is still up for debate, but DeepMind has moved the bar.