Core Input and Output Modes
The model accepts text, images, video, and audio while returning text that can include structured data such as coordinates or bounding boxes.
Token limits reach 131,072 on input and 65,536 on output, which supports multi-turn scene descriptions and longer task sequences.
Breaking Commands into Robot Actions
Natural language instructions are decomposed into ordered subtasks that integrate with robot controllers through function calling or code execution.
Spatial and temporal reasoning lets the model locate objects, track motion across frames, and output 2D coordinates when needed.
- Supported features include search grounding, structured outputs, and a controllable thinking budget for latency versus accuracy tradeoffs.
Privacy Requirements Before Deployment
Operators must obtain consent from any identifiable person before collecting voice, imagery, or likeness data and should apply face blurring where possible.
Review practical privacy steps for consumer devices before running the model on robots that record their surroundings.
Known Limitations in Preview
The API remains in preview, so endpoints and behaviors can change without notice.
Complex queries or high thinking budgets increase latency, and the model can still hallucinate or depend heavily on prompt quality.