About the Talk

Amazon Prime Air delivers packages directly to customers’ yards by autonomous drone, with a click-to-delivery promise of 15 to 60 minutes. The aircraft has no human pilot, not even a remote one, so it must complete each mission and return safely without external help. Part of that is ensuring separation from traffic and obstacles. Prime Air relies on a layered approach in which machine learning based perception, using multiple cameras and radars, is one of the layers. The credibility of that system is what earned Prime Air FAA approval to fly beyond visual line of sight without a pilot, including in uncontrolled airspace.

Amir Navot will describe how machine perception carries that responsibility. Delivery locations are identified without physical markers by matching live video against reference aerial imagery. During descent, segmentation, object detection and stereo depth are fused into a representation that accumulates evidence over time. In transit, the drone detects and avoids other aircraft across full 360-degree coverage, including hot air balloons, paragliders and other traffic that broadcasts no position at all. None of this can depend on the cloud, because no link offers the necessary latency or reliability, so every model runs on the aircraft within hard limits on power, thermal dissipation and weight.

The harder problem is proof. Machine learning at Prime Air is deliberately confined to perception, while action selection and flight control remain conventional and bounded to pre-validated maneuvers. Navot will show how failure budgets are allocated across independent modalities, so that once two modalities are shown to be independent, each needs to clear only a far looser bar. He will explain how training and test coverage is built when the operation itself cannot yet supply data, and why average accuracy metrics conceal failure modes that are rare overall but common in a particular place and season. The recurring lesson: question your model, treat the data and metrics as part of that model, and never let a headline average hide the conditions where a system quietly fails. These methods produce performance predictions that have proven accurate in commercial operation at scale across multiple locations, and that generalize well to new locations where Prime Air never collected data.

Throughout, he will distinguish the methods that generalize to other systems relying on trained models for critical functions from those specific to autonomous drones, and he will close by discussing opportunities to improve the system.