We covered the first reports a few days ago, here's what's actually new now.
Frontier AI that plans, talks to people and controls an entire humanoid body is no longer just a lab demo.
Google DeepMind has released the Gemini Robotics 2 family: a vision-language-action brain that drives whole-body movement, a higher-level embodied-reasoning agent called ER 2, and an on-device variant tuned to run on robots themselves.
ER 2 is the headline change. It accepts interleaved text, images, video and audio, and acts like a planner that understands physical space and can sequence multi-step tasks and talk to people. DeepMind says ER 2 is available to developers through the Gemini API and Google AI Studio, with a private preview on its enterprise agent platform. The full Robotics 2 model and the On-Device 2 variant are currently limited to early-access partners and trusted testers.
Why this matters: up to now, advanced robot reasoning and whole-body control were separate problems. This package stitches them together. That means a single system can plan a task, coordinate legs and arms to move through a space, and adapt when things change, rather than handing off pieces to different subsystems.
How it works, simply: imagine a conductor in front of an orchestra. The conductor hears the room, reads the score and tells each musician when to play. ER 2 is that conductor for a robot, while the Robotics 2 model translates the conductor's plan into precise motor commands. The On-Device model is a compact conductor that lives on the robot itself.
What changes now: robot makers can prototype higher-level autonomy faster, multi-robot coordination becomes easier, and latency or privacy issues fall when parts run on-device. Practical limits remain: you still need humanoid hardware and significant compute to get the full benefits.
What's next: the real test is whether independent teams turn these models into reliable, deployable products outside controlled demos. Will this shift robotics from cloud APIs to owned onboard intelligence? We'll be watching the rollout and the first public benchmarks.