Language and Grounding
Language examples sit after perception and memory. They show how typed observations become captions, grounded references, and small plan payloads without making the runtime depend on a specific LLM provider.
Expected Shape
Section titled “Expected Shape”CaptionObservation summary
scene: red block near blue bowl caption: pick the red block first
GroundingReference resolution
target: "the red one" resolved_object: red_block confidence: 0.92
PlanPrimitive skills
skills: [approach, grasp, lift] status: planned
The point is not a large language model benchmark. The point is the data contract between perception, memory, language, and planning Flows.