Gemini Robotics 2 brings whole body intelligence to robots

TL;DR Highlight

Gemini Robotics 2 controls whole humanoid bodies and enables multi-robot teamwork.

Who Should Read

An API developer wanting to apply LLM agent patterns to robotics and physical AI.

Core Mechanics

Three models split the roles
- Gemini Robotics 2 is a VLA (vision-language-action) model that turns vision and language into motor control.
- ER 2 handles planning and communication, and On-Device 2 is a lightweight VLA that runs locally on the robot.
First whole-body humanoid control
- Earlier models controlled only the upper body for tabletop tasks.
- Apptronik's Apollo 2 walked, picked up objects and placed them on shelves, though movement speed still needs work.
One checkpoint ran three different robots
- The same checkpoint controlled Apollo 2 with SharpaWave hands, Apollo 2 with Inspire hands, and Franka Duo with a Robotiq gripper.
- Whole-body and gripper dexterous tasks reached medium to high success, while multi-finger manipulation remains challenging.
ER 2 acts as the robot's high-level brain
- It observes the room, reasons about steps, hands actions to the VLA and tracks progress.
- It self-corrects when a step fails and handles multi-minute tasks with hundreds of decisions more reliably.

Community Reactions

Opinion that ~60% success is not production-ready
- A commenter cited a separate Google blog with success around 60% and accuracy around 80%.
- They called this page mostly marketing.
Pushback on latency of LLM-driven control
- One commenter argued latency would be 1-2 seconds even on strong GPUs.
- They suggested ML for vision but classic PID-style control for actuation.
Hands-on report that ER 2 is fast
- A VLM Run employee tested it in their visual agent harness.
- Tasks included cropping frames, keyframe extraction, object detection and 3D reconstruction.
Questions about local vs cloud execution
- A commenter said home robots must run locally, not in the cloud.
- Another asked whether stopping near humans means close-contact robots need different control.

How to Apply

Test ER 2 in AI Studio with a multimodal task
- Open Gemini Robotics ER 2 in Google AI Studio today and upload a task video or image.
- Ask for step-by-step plans and for when the task starts and ends.
Split planner and executor in your agent design
- Let the top-level LLM only plan, track progress and retry on failure, while a separate module executes.
- Have the planner verify each step's success.
Add unsafe tool-call refusals to your eval set
- Build test cases where the agent receives risky tool calls, following the ASIMOV-Agentic idea.
- Also verify that it escalates to a human when uncertain.

Try It

copy & run

# Paste into Claude or ChatGPT (simplified version of the method: planner verifies executor)

You are a task planner. Break the task below into executable steps and write a success check for each.
Task: {{task description}}
Available tools: {{tool list}}

Rules: 1) Refuse steps that look unsafe or impossible and explain why. 2) Ask a human when uncertain. 3) If a step fails, propose an alternative plan.

Next: mix in deliberately unsafe tool calls and check that it refuses them.

Terminology

VLAA model that takes vision and language input and outputs motor actions. Think of it as a translator from instructions to muscle signals.

ER (Embodied Reasoning)A reasoning model that understands the physical world and plans multi-step tasks. It acts like an on-site supervisor.

embodimentA robot's physical configuration (arms, hands, sensors, degrees of freedom). It is why moving software between robots is hard.

end effectorThe hand or gripper at the end of a robot arm.