Google DeepMind Unveils Gemini Robotics-ER 1.5: Smarter Robots

Google DeepMind has taken a major step toward making robots more capable and reliable in the real world.

The company announced Gemini Robotics-ER 1.5, an advanced “thinking” AI model that allows robots to better understand their surroundings, plan multi-step tasks, and even use tools like web search or robot APIs to complete complex jobs.

What Was Announced

DeepMind introduced two new models:

Gemini Robotics-ER 1.5 – a high-level reasoning model that analyzes environments, creates task plans, and calls external tools when needed.

Gemini Robotics 1.5 – a vision-language-action (VLA) system that converts those plans into real robot motions.

The ER 1.5 model is now broadly available to developers through the Gemini API in Google AI Studio.

It is labeled as gemini-robotics-er-1.5-preview and supports long-context, multimodal inputs such as text, images, video, and audio.

The system’s knowledge was last updated in September 2025 with a cutoff of January 2025.

Meanwhile, the action model Gemini Robotics 1.5 is rolling out to select partners.

Key Features

Embodied reasoning: ER 1.5 specializes in visual and spatial understanding, task planning, progress estimation, and tool use, enabling robots to “think before acting” and explain intermediate steps for transparency.

Spatial grounding: The model returns structured outputs like 2D points or bounding boxes to locate objects, which can be fused with a robot’s 3D sensors to generate accurate motion plans for manipulation tasks.

Long-context, multimodal input: ER 1.5 ingests images, video, audio, and language to decode scenes and instructions, with developer docs outlining supported data types and token limits for extended planning and perception.

Tool calling and orchestration: The model can natively call tools such as Google Search for local knowledge and can invoke user-defined functions or a VLA model to execute steps in the physical environment.

Thinking budget: Developers can tune a “thinking budget” to trade latency for accuracy—quick budgets for spatial tasks like detection, and larger budgets for complex reasoning like counting or estimating weights.

Learning Across Robots: Gemini Robotics 1.5 can share skills between different robots. A task learned on one robot can work on others like ALOHA 2, Apollo humanoid, or Franka arms.

This makes training faster and lets AI be used on many types of robots.

Technical Notes: The ER 1.5 preview can handle multiple inputs (text, images, video, audio) and give structured outputs for precise actions.

Developers can set a “thinking budget” to choose between faster or deeper reasoning.

Google also offers a Robotics Cookbook and safety rules to help teams use the system responsibly.

Availability and pricing

Gemini Robotics-ER 1.5 is live now for developers through Google AI Studio and the Gemini API.

Gemini Robotics 1.5 is limited to select partners.

Pricing includes a free tier and a paid tier, with charges based on tokens processed. Some limits apply for grounded search requests, depending on region and usage.

News Gist

Google DeepMind has launched Gemini Robotics-ER 1.5, an AI model that helps robots “think before acting.”

It improves planning, tool use, and safety, enabling multi-step tasks like sorting, packing, and cleaning.

Developers can now access ER 1.5 via Gemini API.

FAQs

Q1. What is Gemini Robotics-ER 1.5?

It’s a new AI model from Google DeepMind that helps robots reason, plan, and complete multi-step real-world tasks.

Q2. How is it different from Gemini Robotics 1.5?

ER 1.5 handles reasoning and planning, while Robotics 1.5 executes physical actions.

Q3. Who can use Gemini Robotics-ER 1.5?

It’s available to developers through Google AI Studio and the Gemini API.

Q4. What are example tasks it can do?

Sorting recycling, packing items, or cleaning up spaces by breaking instructions into smaller steps.

Q5. Is it safe?

Yes. DeepMind uses multi-layer safety checks, including collision avoidance and updated safety benchmarks.