Google DeepMind Ships Gemini Robotics 2, Giving Humanoids Whole-Body Control and Five-Finger Dexterity
Google DeepMind released a three-model suite — Gemini Robotics 2, ER 2, and On-Device 2 — that moves robots past tabletop manipulation into full-body locomotion and multi-robot teamwork.
Google DeepMind released Gemini Robotics 2 on July 30, a three-model suite that pushes its robotics stack past tabletop pick-and-place tasks into full-body humanoid control. The lineup: Gemini Robotics 2, a vision-language-action (VLA) model that converts vision and language input directly into motor commands; Gemini Robotics ER 2, an embodied-reasoning model for multi-step planning and multi-robot coordination; and Gemini Robotics On-Device 2, a compact VLA that runs locally and adapts to a new robot body with just a few hours of data.
The headline capability is whole-body control — walking, crouching, and reaching in a single coordinated motion — paired with five-fingered dexterity that lets a humanoid pick up irregular objects the way earlier two-finger grippers couldn’t. ER 2 adds the planning layer: it can break a task into steps, communicate intent to a human collaborator, and hand off subtasks across multiple robots working the same physical space.
Google lined up hardware partners to show the range: Apptronik’s Apollo 2 humanoid, Franka’s Duo dual-arm platform, Boston Dynamics, and Agile Robots. That’s a deliberate signal that Gemini Robotics is meant to be embodiment-agnostic — the same reasoning stack driving a warehouse humanoid and a factory-floor arm.
Access is staged. ER 2 is available now in Google AI Studio and in private preview on the Gemini Enterprise Agent Platform — the easiest on-ramp for developers who want to prototype planning and reasoning without touching a physical robot. The full VLA and On-Device models stay gated behind an early-access partner program, which is standard practice for anything that outputs motor torque commands to real hardware.
DeepMind also introduced ASIMOV-Agentic, a new safety benchmark specifically for verifying agentic behavior in physical robots — a tacit admission that a model choosing its own multi-step physical actions needs a different bar than one generating text. That’s notable timing: it lands the same week more than a thousand AI-lab employees, including several from DeepMind, signed a letter asking governments to build tools for pacing frontier AI development before autonomy outruns oversight.
Competitively, this is Google pressing an advantage. Figure, 1X, and Tesla’s Optimus program are all racing toward general-purpose humanoids, but the hard part has never been walking — it’s coordinating perception, reasoning, and fine motor control in one loop fast enough to be useful. Shipping ER 2 broadly while keeping the VLA gated lets Google collect real deployment data from partners before opening the riskiest layer to the public.