Google DeepMind Launches Gemini Robotics 2 Family
On July 30, 2026, Google DeepMind launched the Gemini Robotics 2 family—whole-body robot control (Gemini Robotics 2), high-level embodied reasoning and multi-robot collaboration (Gemini Robotics ER 2), and an on-device adaptation tier—extending Gemini from tabletop manipulation to full humanoid locomotion and multi-robot teamwork.
TLDR
Google DeepMind on July 30, 2026 launched the Gemini Robotics 2 family: Gemini Robotics 2 (whole-body motor control), Gemini Robotics ER 2 (embodied reasoning “high-level brain” for planning, tool use, and multi-robot collaboration), and an on-device adaptation tier. Demos include Apptronik Apollo 2 packing, kitchen chores, and multi-robot garage cleanup. Bloomberg and Ars Technica same-day coverage confirm full-body humanoid control beyond prior upper-body-focused stacks; at least one ER-facing path is described as developer-available.
What shipped
| Component | Role (lab + secondary primaries) |
|---|---|
| Gemini Robotics 2 | Whole-body VLA-style control: locomotion + manipulation (walk, crouch, grasp, carry) |
| Gemini Robotics ER 2 | High-level planner: chat, spatial reasoning, multi-step tasks, handoff to lower-level VLA; can call tools (e.g. Search) |
| On-Device 2 | Edge adaptation; secondary reports claim few-hour / <200-example body transfer |
| Multi-robot | Different robot morphologies collaborating on shared tasks |
| Hardware demos | Apptronik Apollo 2, Duo robots, fine motor (knot-tying, lightbulb) |
Google Labs/DeepMind posts frame ER 2 as the brain that thinks about the next step while acting, and Robotics 2 as the body stack that extends past tabletop-only Gemini Robotics generations.
Access and competitive context
Public developer access is partial: Ars Technica notes a trio of models with one publicly available at launch via Gemini API robotics docs; action/on-device tiers remain early access in several secondary writeups. Competitive set: Figure, Tesla Optimus stacks, OpenAI/Microsoft robotics research, Chinese humanoid labs—DeepMind’s differentiator is Gemini-scale reasoning + multi-robot orchestration under one brand.
Why this story matters
Physical AI is the next frontier product surface after chat and coding agents. A named Robotics 2 family with whole-body control and multi-robot collab is equal-strength lab primary: not a research paper alone, but a productized model line with partner hardware demos. Watch real-world failure rates, safety stops, transfer across robot bodies, and how much of the stack is API-open vs partner-gated.