What Happened
Google DeepMind announced the Gemini Robotics 2 family on July 30. Gemini Robotics ER 2 became available through Google AI Studio, with an enterprise private preview. The action and on-device models were offered to early-access partners.
Why It Matters
The family extends DeepMind's earlier upper-body manipulation systems toward coordinated humanoid movement, fine manipulation and collaboration between robots. It connects physical reasoning with action across multiple embodiments. The demonstrations remain bounded: DeepMind explicitly reports continuing difficulties with multifinger dexterity and movement speed.
Technical Details
Gemini Robotics 2 converts vision and language into motor actions. ER 2, based on Gemini 3.5 Flash, supplies planning, progress tracking and tool orchestration. The separate On-Device 2 model runs locally; DeepMind reports adaptation to new bi-arm platforms using typically fewer than 200 examples. Its model card limits the evaluation scope primarily to standing bi-arm manipulation, rather than whole-body mobility.