Smoother streams, faster responses, less repetition
Depth cameras stream 47x smaller, speaker audio no longer stutters, the AI responds faster, and the robot won't repeat itself after speaking.
A Collection of Quality-of-Life Improvements
This update bundles several optimizations that make daily use noticeably smoother.
Depth Camera Streams Without Lag
Depth cameras now use zlib compression (image/depth-zlib format), reducing bandwidth by 47x. Even on slow or congested networks, depth visualization stays smooth and responsive in the dashboard.
G1 Speaker No Longer Stutters
On Unitree G1, the speaker process is now isolated from heavy sensor traffic (like lidar point clouds). Audio playback stays smooth even during intensive perception tasks — no more choppy speech during navigation.
Robot Won't Repeat Itself
Previously, after the robot finished speaking, the TTS completion event sometimes triggered extra AI reasoning rounds — causing the robot to say the same thing 2-3 times. Fixed: speech completions no longer trigger unnecessary LLM invocations.
Dashboard Shows Model & Response Time
The activity log now displays which LLM model is responding and how long each round takes (e.g. "glm-5.2 4.7s"). Useful for debugging latency or comparing model performance.
Faster AI Responses
Event text fed into the LLM has been slimmed down — redundant fields stripped, empty monitoring cycles skipped, duplicate ASR events deduplicated. The result: less token waste and slightly faster responses across all robots.