General-purpose robots need to understand object relationships and predict the consequences of actions — not just detect what's in view. The current state of the art is a fragmented stack: YOLO for detection, SAM for segmentation, custom post-processing for relationship inference, and handwritten glue that breaks every time the environment changes. Each component adds latency, adds a failure mode, and adds a training budget.
Darwin replaces that stack with a single API call. It runs on Jetson-grade edge hardware with 45ms p50 per-frame inference, produces deterministic outputs with calibrated confidence scores, and works zero-shot across environments. The same model that understands a warehouse picking task understands a kitchen manipulation task — because it learned to understand video, not to detect a specific set of objects.
Robotics integration guide and edge deployment examples — coming during our August launch.