Gemini Robotics 2: Whole-Body Control Is the Product Line—API Access Is Partial
Google DeepMind launched Gemini Robotics 2 (whole-body), ER 2 (embodied reasoning and multi-robot collab), and an on-device tier—demos on Apptronik Apollo 2. Differentiator is Gemini-scale planning plus multi-robot orchestration; public developer access is still only partial at launch.
Corrected October 14, 2026: Attributed the on-device adaptation figures (a few hours of data, or ~200 examples) to Ars Technica's report, and attributed full-body-control claims to Google as reported.
Physical AI is the next frontier product surface after chat and coding agents. DeepMind’s stake is a named Robotics 2 family—whole-body control plus multi-robot orchestration under one Gemini brand—not another tabletop-only research paper.
Google DeepMind launched 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 report Google’s claims of 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; Ars Technica reports adaptation to new robot designs with a few hours of data, or ~200 examples |
| 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 vs competitive set#
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.
Limits#
- Real-world failure rates and safety-stop behavior are demo-led, not independently scored here.
- On-device transfer claims (few hours / ~200 examples) are secondary reports, not a DeepMind audited table.
- Most of the stack remains partner-gated vs fully API-open.