Google DeepMind debuts Gemini Robotics 2: full-body control, multi-robot coordination, safety benchmark

Almost a year has passed since version 1.5 — an eternity in this field — and Google DeepMind is releasing Gemini Robotics 2. The main leap isn't in some synthetic benchmark but in anatomy: the underlying VLA (Vision-Language-Action) model controls a full humanoid body for the first time, from legs to fingertips. The hands come with 22 degrees of freedom, but let's be precise: this is still not a surgical robot. Success rates on multi-finger tasks hover around 30 to 40 percent. That's far from perfect, but compared with what was available until now, it's a step forward that holds water.

Sitting above the VLA is Gemini Robotics ER 2, a separate model that is essentially a classic VLM. Where the VLA issues low-level motor commands, the ER handles the reasoning grunt work: breaking a task into subtasks, identifying objects and coordinates, and the big novelty, coordinating several heterogeneous robots for a single job. In addition, the ER can intervene in the VLA's actions at any stage, a critical capability for self-correction when things go wrong mid-run.

On the local deployment front comes Gemini Robotics On-Device 2, a distilled version of the VLA for edge inference. Its practical advantage is adaptation speed: because the whole family was trained on heterogeneous multi-embodiment data, you can take this model, run about 200 demonstrations for a few hours, and get reasonable performance on an entirely new body with custom sensors, kinematics and DoF. No heavy fine-tuning, no months of integration. That's the part that turns the promise of "general robotics" into something you can pilot tomorrow morning.

For dessert, Google launched ASIMOV-Agentic, the first benchmark of its kind for evaluating the safety of robotic agents. Naturally enough, their own model was declared the "world's safest" according to the test they wrote themselves. The tongue-in-cheek nature is obvious, and the community will no doubt pick apart the methodology, but the mere existence of an agreed-upon metric set — even if it's one-sided for now — is necessary progress for a field that still operates a bit like the Wild West.