The tutorial, published on the DeepCausality website with its source code in the project repository, builds a drone fail-safe that adapts its response while a failure unfolds.
Why a dynamic fail-safe
A conventional drone fail-safe fires on a single condition, such as a lost position fix or a crossed geofence. It ignores what lies below the drone and who stands near it. The tutorial starts from two show-drone accidents, Melbourne in 2023 and Orlando in 2024, where each drone's fail-safe worked this way.
The tutorial
In the tutorial, a drone inspects a high-voltage line above a grass slope with a creek, trees, a ravine, and a work crew below. Mid-flight, the drone loses GPS, then its command link, then a battery cell. Each of the tutorial's five parts adds one layer of DeepCausality. Dynamic causality confirms the faults; dynamic context reads the ground through a thermal camera and LiDAR; and a causal state machine flies the drone to the nearest safe patch. The fourth part adds the Effect Ethos, which turns published emergency procedures for drone pilots into norms that rule on every maneuver before the drone flies it.
Results
The fifth part flies the textbook, context-and-action, and Effect Ethos versions through the same 1,000 randomized scenarios. A touchdown within 10 m of a person counts as putting that person at risk. The textbook fail-safe put a person at risk in 3.5 % of scenarios; the Effect Ethos version did so in 0.1 %, with a one-sided 95 % upper bound of 0.47 %. The Effect Ethos version loses more drones than the context-and-action version, 36.3 % against 26.8 %, because it ranks people above the aircraft. The campaign also caught six failures in earlier versions of the protocol before deployment. These rates describe the tutorial's simulated world.
Provenance and governance
The Center contributes its verified work back to DeepCausality, an open-governance project of the Linux Foundation for Data & AI. The tutorial exercises the full DeepCausality stack, from confirming a fault to ruling on an action, in one worked application.