Gemini Robotics 2 brings whole body intelligence to robots

Google DeepMind introduces three models: a VLA for whole-body and dexterous control, ER 2 for multi-minute planning and collaboration, and On-Device 2 for local inference. The VLA is shown on Apollo 2 humanoids and a Franka Duo, including walking, knot-tying and packing.

On-Device 2 is described as adapting to new bi-arm bodies in a few hours, typically with fewer than 200 examples. ER 2 is public in AI Studio; the VLA and on-device models are early-access only.

This is Google DeepMind’s own launch post, so success rates, adaptation times and safety claims are vendor-reported. The post itself says multi-finger dexterity remains challenging and movement is still slow.

Signal Desk's read: the product is a stack, not a single robot brain. Whole-body control is the novelty relative to table-top upper-body work, but usefulness still hinges on speed, reliability and partner hardware.

Fast on-device adaptation is the more commercially interesting claim. If a few hours and a few hundred examples really retarget a body, the bottleneck shifts from training a new policy to collecting clean demonstration data.

The post introduces ASIMOV-Agentic, a benchmark for refusing unsafe tool calls and requesting help under uncertainty. Personal-safety claims include stopping when a person approaches.