A robot that can make a bed only in the room where it was trained has limited generalization. A system that can enter a new home and perform the learned behavior without collecting training data there demonstrates a stronger form of transfer.
Does zero-shot mean the robot was never trained?
No. The model can be extensively pretrained and a behavior can be specified using training data. “Zero-shot” describes the target evaluation environment or objects: no environment-specific fine-tuning is performed before deployment there.
Figure Helix 2.5 example
Figure reported Helix 2.5 performing room tidying, towel folding and bed making across 30 unseen homes without collecting data or fine-tuning in those homes. Figure says pretraining on its Index human-behavior dataset increased zero-shot success from 9% to 56% in its controlled comparison.
Why is generalization important?
Robots cannot economically be retrained for every factory aisle, warehouse, home and object. Generalization is therefore central to turning robot foundation models into systems that can be deployed across many real environments.
Zero-shot vs few-shot adaptation
Zero-shot deployment uses the existing model without target-environment adaptation. Few-shot or rapid adaptation allows a small amount of new data. Google’s Gemini Robotics On-Device 2, for example, emphasizes rapid adaptation to new robot embodiments with only a few hours of data—a related but different capability.
Related terms
Sources
Physical-AI terminology is evolving quickly. We use the broader technical meaning and distinguish it from individual vendors’ model names and claims.