Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

Today's autopilots are designed to be as simple as possible.

The intention is that a pilot should be able to fully understand the complete autopilot algorithm and understand exactly how it should act in any situation.

Today's autopilot can't handle any emergencies at all, they are designed to disconnect when they detect something weird. They don't even understand aerodynamics.

I don't know where Airbus' new system lies, but if we actually designed a fully autonomous plane using modern traditional AI techniques (so no machine learning or neural networks; Just decision trees, adaptive aerodynamic models and predictions) then it could handle a wide range of emergencies.

It should be able to handle bad weather and engine out emergencies just fine, and should even be able to land in the Hudson River or anywhere else it has detailed maps.

These are all documented flight emergencies that any programmer of a fully autonomous plane would know about and test their software against in simulations.

The main areas where such fully autonomous planes would have issues are avionics failures, misleading sensors, emergency landings in areas with substandard mapping and unexpected aerodynamic failures (beyond what their adaptive aerodynamic model can handle)



Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: