How we built a desktop companion robot with Gemma 4 and Raspberry Pi

A Google AI team built DinoDesk AI, a LEGO-bodied desk robot running a hybrid local/cloud architecture. Gemma 4, running locally through LM Studio, handles routine chat with near-instant response and full privacy, since the robot has no camera. Gemini Flash takes over for heavier reasoning, coding help, or complex explanations, with a router on the user’s PC deciding which engine answers each request. A five-state finite state machine coordinates the robot’s LCD eyes, 8-bit sound, and motor movements so the character reads as alive rather than assembled from separate parts.

Fichamento

#

• A Raspberry Pi on the desk sends identical requests over Wi-Fi to a Hybrid LLM Gateway on the user’s PC, regardless of which engine ends up answering.
• Local mode, Gemma 4 via LM Studio, delivers about 200ms first-token latency for casual chat, fully offline and private. Cloud mode, Gemini Flash, takes over for coding help or deep explanations at roughly 800ms first token.
• Auto-Hybrid mode defaults to the local model and escalates to the cloud only when a prompt shows multi-step reasoning keywords like “explain,” “compare,” or “write code,” visible to the user as the status light shifting from green to gold.
• The body is built from LEGO bricks and Technic mechanisms instead of a sealed 3D-printed shell, so it can be customized or repaired by hand.
• A five-state finite state machine, Sleeping, Idle, Listening, Thinking, Speaking, coordinates the LCD eye expressions, 8-bit sound cues, and neck and tail motor movement for each state.
• Four hardware buttons handle cancel and mute, servo re-centering, expression and engine switching, and a tail-wag test with volume control, alongside voice and sensor triggers.