A grounded language model links language behavior to physical experience or verified knowledge. Sensors such as cameras and microphones let an embodied system compare language-based assumptions with the surrounding world. section 10.2.3
Grounding addresses a central weakness of text-only language models: fluent output can still contain nonsense because internet text provides no direct opportunity to detect mistakes against reality. Embodied feedback can expose faulty common-sense rules and support correction. section 10.2.3
Counting apples
One apple plus another apple produces two apples in the physical world. A grounded model could use that observable result to check its language-based reasoning and improve counting or addition. section 10.2.3
