Google’ AI Robots: Slam Dunks and Safety Smarts

Generative AI is stepping into the real world, and Google DeepMind’s latest models are teaching robots to dunk basketballs and fold origami. Built on the versatile Gemini foundation, these AI agents promise smarter, safer robots that adapt on the fly, bringing us closer to a future of helpful machines.

At a Glance

• Google DeepMind unveils Gemini Robotics and Gemini Robotics-ER.
• Robots handle tasks like dunking a basketball and adapting to changes.
• Focus on generalization, adaptability, and dexterity for real-world use.
• Safety features and partnerships aim to shape tomorrow’s robots.

What’s New with Gemini Robotics?

Smarter Moves, Bigger Wins

Gemini Robotics blends vision, language, and action to guide robots through tasks they’ve never seen. In a demo, a robot arm dunks a mini basketball after a simple voice command, “Pick up the basketball and slam dunk it.” “It’s never seen basketball before,” says Kanishka Rao, principal software engineer, in a briefing. Yet, it connects the dots using Gemini’s broad knowledge, as shared in IEEE Spectrum read more here.

Adapting Like a Pro

Robots need to roll with the punches. Carolina Parada, head of robotics, highlights adaptability in a demo where a robot tracks a moving container to place plastic grapes. “It handles changing instructions and circumstances better than past models,” she says. This flexibility is key for robots to thrive outside controlled labs.

Embodied Reasoning: Thinking Like Us

Grasping the World

Gemini Robotics-ER adds “embodied reasoning,” letting robots figure out how to interact with new objects. Parada notes it picks a coffee mug by the handle, mimicking human habits. But here’s the catch: human-style grasping might not always suit a robot’s strengths, like gripping a hot mug directly.

Beyond the Basics

This reasoning isn’t just for show. It’s about intuition, like guessing how to stack odd-shaped blocks. DeepMind’s goal? Robots that don’t just follow scripts but think through real-world puzzles, making them more useful every day.

Safety First, Always

Layers of Protection

Safety’s no joke when robots move in our space. Vikas Sindhwani, head of robotic safety, explains a multi-layered approach: physical controls stop crashes, while “semantic safety” checks if actions make sense. Gemini Robotics-ER even weighs risks, like avoiding a hot stove.

Sharing the Safety Love

DeepMind’s not hoarding their safety tricks. They’re releasing the Asimov benchmark to test common-sense smarts, with Gemini scoring over 80 percent. “Safety isn’t competitive,” Sindhwani says. Check out the details in their technical report.

Robots on the Horizon

Teamwork Makes the Dream Work

DeepMind’s partnering with Apptronik to build Gemini-powered humanoid robots. Plus, elite testers like Boston Dynamics are getting early access. These collabs could soon put advanced robots in homes and workplaces.

Conclusion

Google DeepMind’s Gemini Robotics models are a slam dunk for AI, blending smarts, adaptability, and safety into robots that learn like us. From basketball nets to coffee mugs, they’re paving the way for a robotic revolution.

Want more AI insights? Subscribe below and explore AI trends at ainewzworld!