Gemini Robotics 2 brings whole body intelligence to robots

What it claims

Google DeepMind's Gemini Robotics 2 release has three models. Only one of them is usable without a partnership. Gemini Robotics ER 2, the embodied-reasoning VLM, is available on Google AI Studio and in private preview on Gemini Enterprise Agent Platform. The Gemini Robotics 2 VLA and Gemini Robotics On-Device 2 go to early-access partners only. The page gives no benchmark numbers for any capability claim. Everything is shown through demos and qualitative statements.

How it works

The models split the work into a planner and an executor:

• ER 2 is the high-level agent. It takes the user instruction, observes the scene, plans the steps, and calls the VLA to carry out actions. It tracks progress until the task is done, and it retries or self-corrects when a step fails.
• ER 2 tracks task timing. It detects when tasks begin and end and pinpoints when key events happen. The page calls this a step change in progress understanding. Target tasks last "several minutes" and involve "hundreds of decisions".
• ER 2 also handles safety. It can refuse unsafe tool calls coming from the VLA, predict whether a task is feasible, and ask for human help when it is uncertain. When it detects a person too close, it triggers a safety tool call that brings the robot to a safe stop.
• Multi-robot collaboration. Robots of different types communicate to split up a workflow. The page does not describe the mechanism.
• On-Device 2 adapts to new robot bodies. It is natively multi-embodiment and inherits the "motion transfer" technique from Gemini Robotics 1.5. It adapts to a new bi-arm embodiment in a few hours, typically with fewer than 200 examples, including bodies with different shapes, sensors and degrees of freedom. The demonstrated platforms are Dexmate, SO101 and Trossen.

flowchart TD
  U[User instruction] --> ER[ER 2: observe, plan]
  ER -->|feasible?| Q{Confident?}
  Q -->|no| H[Request human intervention]
  Q -->|yes| V[VLA executes step]
  V -->|tool call| S{Safe?}
  S -->|no| R[ER refuses / safe stop]
  S -->|yes| P[Track progress]
  P -->|step failed| ER
  P -->|next step| ER

Evidence

All evidence on the page is demonstrations:

• Whole-body humanoid control on Apptronik Apollo 2. In the example, the robot walks to a table, picks up a watering can, and places it in a bin on a bottom shelf.
• Dexterity with the 22-DOF SharpaWave hand: tying knots and sealing a ziplock bag.
• Tight packing with parallel grippers on Franka Duo.

ER 2 is described as the "safest robotics model to date" on safety-constraint-following and human-proximity benchmarks. No scores or baselines are given. A separate Gemini Robotics 2: Safety Technical Report is referenced for details.

How to use it

• Try ER 2 now: it is on Google AI Studio.
• Enterprise use: apply for the Gemini Enterprise Agent Platform private preview.
• VLA and On-Device 2: these require early-access partner status.
• Integration: the page points to a separate Developer blog post on bringing the models to your own hardware.

The page states no API shape, model sizes, pricing or licence terms.

Limits

• Speed and precision. The authors say movement speed still needs work, and that precision and speed fall short of human-level dexterity.
• Embodiment coverage. Fast adaptation is claimed for bi-arm embodiments only. Whole-body control is shown only on Apollo 2.
• No quantitative results. Success rates, the ASIMOV-Agentic results, and the adaptation claim ("few hours, <200 examples") come with no metrics or baselines. Check the Safety Technical Report before relying on any safety claims.