Google DeepMind has released Gemini Robotics—putting large language models into robots.
Demo Scenarios
At the launch event, researchers told the robot, "Help me clean up the table, throw away the trash, and put the cup in the dishwasher," and the robot autonomously completed: - Identifying different objects on the table (food scraps vs. cups) - Planning action paths - Executing fine-grained manipulation (picking up a fragile cup) - Handling anomalies (when the trash can was full, it switched to another one)
The entire process required no pre-programmed routines, relying entirely on the semantic understanding and reasoning capabilities of the large model.
Why This Matters
Previously, robots needed to be individually programmed for each task. With large models, robots can "understand" natural language instructions and adapt to new scenarios.
This is essentially giving robots a "brain."
Challenges
- Safety: If an AI-controlled robot makes a mistake, the consequences can be severe
- Reliability: Large models can sometimes hallucinate
- Speed: Real-time physical interaction demands extremely low latency
Industry Impact
Manufacturing, logistics, and domestic services are likely to be transformed first. Google says they will initially collaborate with logistics companies on warehouse robots.
The AI-ification of the physical world may happen faster than we think.
References:
Comments
Comments are closed.