Gemini is now DeepMind's office tour guide

Remember Everyday Robots, the ambitious Google hardware project that got unceremoniously shut down during last year's layoffs? It turns out those wheeled machines didn't just end up gathering dust in a storage closet.

This week, Google DeepMind's Robotics team brought them back to life. With a jaunty yellow bowtie attached and Gemini 1.5 Pro serving as their onboard brain, these robots are now navigating a 9,000-square-foot office space by simply listening to people and reading whiteboards.

Office tours and whiteboard maps

In a new paper detailing an architecture called Mobility VLA, DeepMind showed off how long-context vision-language models can turn an office tour into an actionable navigation system.

Instead of programming rigid CAD maps into the machine, researchers used a technique called MINT, or Multimodal Instruction Navigation with demonstration Tours. In plain English, an engineer literally walks the robot around the office once, pointing out landmarks with regular speech. Gemini ingests the tour, blends spatial layouts with common-sense reasoning, and creates a mental map of the space.

Once familiarized, the robot responds to natural speech triggered by a smart-assistant-style "OK, Robot" wake word. In one demo, an employee asks the robot to take him somewhere he can draw things. The machine pauses, says it is thinking with Gemini, and promptly guides him to a massive office whiteboard.

Even more impressively, another employee tells the robot to follow directions sketched directly on that whiteboard. Staring at a crude hand-drawn map pointing toward a testing space called the Blue Area, the robot calculates a route and navigates right to it before proudly announcing its own success. Google claims the system hit an approximate 90% success rate across more than 50 real interactions.

Why it matters

We think this project demonstrates one of the smartest shortcuts in robotics right now: stop obsessing over manual route-planning algorithms and just hand the high-level perception over to an existing frontier model.

Traditional navigation systems choke on ambiguity. If you tell a regular automated cart to take you somewhere to draw, it fails because there is no room officially named drawing. Gemini 1.5 Pro bridges that gap with common-sense world knowledge, deducing that drawing implies a whiteboard, while reading scribbled hallway diagrams that would baffle standard computer vision.

That said, the cracks still show. The robot openly admits it needs about a minute to think before moving, a latency window that works fine for research videos but falls flat in fast-paced retail or hospital logistics. Fifty trials is also a modest sample size, and a 10% failure rate in an office is still an issue if a machine gets lost once every ten errands.

Still, it is a fascinating proof of concept. Google effectively showed that abandoned robotics hardware can instantly gain new commercial life simply by plugging into long-context multimodality, nudging embodied AI closer to practical reality.