Google DeepMind today unveiled Gemini Robotics On-Device, a language-first model that runs entirely on a robot’s internal hardware, erasing the latency and privacy concerns of cloud dependence. The upgrade builds on the Gemini release from March but compresses inference to fit on ARM-based boards, allowing factory bots and home helpers alike to understand nuanced voice prompts, identify unfamiliar objects and adapt actions in real time—even without Wi-Fi.

Engineers claim a 35 percent speed gain in visual-motor tasks and a 60 percent drop in power consumption compared with its cloud-tethered predecessor. Developers can fine-tune behaviours via text instructions, then transfer custom weights over USB—no retraining farm required. Analysts say the move could accelerate adoption in healthcare logistics, where patient-data sensitivity limits external connections. Still, the breakthrough arrives amid cautionary forecasts: a Gartner note released hours later predicts 40 percent of “agentic AI” initiatives will be scrapped by 2027 for failing safety audits and ROI tests. DeepMind counters that running locally strengthens controllability because state changes are logged on-device, simplifying certification. Early prototypes are headed to Boston Dynamics and Japan’s Mujin for pilot deployments this quarter.