Avride adds cloud vision-language models as safety layer for delivery robots

Avride Inc. has integrated cloud-based vision-language models into its sidewalk delivery robots as an automated early-warning system. The VLM-watcher supplements the robots' onboard perception stack by providing holistic scene understanding. It evaluates camera snapshots sent every few seconds to flag sensitive situations such as active crime scenes or unmapped roadwork, allowing remote operators to intervene before the robot enters restricted areas. Visual data is anonymized on the robot before transmission. The company says the approach significantly reduces the chance of robots entering emergency zones or crossing live construction sites while maintaining local autonomy for navigation.