Google's New AI Robots Learn to Think and Adapt Like Humans

Google DeepMind has unveiled two groundbreaking AI models—Gemini Robotics 1.5 and Gemini Robotics-ER 1.5—that significantly enhance robots' ability to reason, plan, and adapt autonomously. These models represent a shift from single-task robots to those capable of multi-step problem-solving, marking a major milestone toward Artificial General Intelligence (AGI) in the physical world.

Key Advancements

• Generalization: Robots can now apply knowledge to new scenarios, overcoming a longstanding hurdle in robotics. For example, they can sort laundry by color, pack suitcases based on weather forecasts, or check local recycling rules—tasks requiring contextual awareness.
• Online Integration: The robots can access Google Search for real-time information, such as recycling guidelines in San Francisco, and execute tasks accordingly.
• Collaborative Learning: Skills learned by one robot can be transferred to others, accelerating adaptability.

Performance and Limitations

In tests, the robots demonstrated a 20-40% success rate in executing complex tasks—a notable improvement, though far from perfect. For instance:

• They visually identified objects, cross-referenced them with online recycling rules, and disposed of them correctly.
• They employed chain-of-thought reasoning, breaking tasks into logical steps (e.g., "whites go in one bin, colors in another").

Technical Breakdown

The two models work in tandem:

1. Gemini Robotics-ER 1.5: Acts as the "brain," planning tasks and retrieving information via search (e.g., weather forecasts). It’s a Vision-Language Model (VLM).
2. Gemini Robotics 1.5: Handles physical execution, converting instructions into movements. It’s a Vision-Language-Action (VLA) model.

Google CEO Sundar Pichai touted the models as "the next big step toward general-purpose robots."

Industry Context

Google’s approach contrasts with competitors:

• Tesla: Focused on mass-producing factory robots.
• Boston Dynamics: Prioritizes physical agility (e.g., backflipping Atlas).

Meanwhile, China leads in industrial robotics, with 1.8 million operational robots in 2023 (International Federation of Robotics).

Availability

• Gemini Robotics-ER 1.5: Now accessible via Gemini API.
• Gemini Robotics 1.5: Limited to select partners.

This development underscores the accelerating global race toward intelligent, adaptive robotics.