Model inputs, outputs, and limits
The model accepts text, images, video, and audio and returns text. Maximum input is 131072 tokens and maximum output is 65536 tokens.
It is currently labeled preview with model code gemini-robotics-er-1.6-preview. The latest update was December 2025 and knowledge cutoff is January 2025.
Core robotics capabilities
The model performs object and scene understanding, spatial and temporal reasoning, and produces structured output such as points or bounding boxes.
It can break a high-level command into subtasks, call functions, and run code to orchestrate longer sequences of actions.
Thinking budget for latency versus accuracy
A flexible thinking budget lets developers trade speed for depth. Smaller budgets work well for quick spatial tasks such as object detection.
Larger budgets improve results on counting, weight estimation, or other multi-step reasoning problems but increase response time.
Safety and consent requirements
Users must obtain consent before collecting voice, imagery, or likeness of identifiable people. The operator remains responsible for maintaining a safe physical environment around the robot.
Model hallucinations can still occur on ambiguous inputs, so every action that moves hardware needs external checks.
Integration patterns and current limits
Prompt quality strongly affects results. Clear instructions that specify coordinate systems and success criteria reduce errors on physical tasks.
Preview status means API details may change. Complex queries or high thinking budgets add latency, so test end-to-end timing on target hardware before deployment.