Musk on Tesla Autopilot vs aircraft autopilot
“It would no more make sense to disable Tesla’s Autopilot… than it would to disable autopilot in aircraft…”
This talk is dangerous.
The sky at 40,000 feet almost never contians external objects. No concrete barriers, no tractor-trailers with white sides, no stationary firetrucks attending accidents, no stop signs, no temporary road detours, no manual traffic direction, none of the gazillion other germane objects and situations that occur in the road environment. At 40,000 feet, it’s just air.
Hence, aircraft autopilot systems don’t try to “recognize” external objects. Aircraft autopilots don’t have lidar, sonar, radar, video cameras hooked up to the autopilot systems. Aircraft autopilots don’t use AI software based on deep learning algorithms run against vast data sets. They don’t use AI software, because aircraft flight environment is air.
The aircraft system that detects other planes is based on receiving a simple radio signal, not object identification. Ground proximity warning is a separate function and doesn’t control the aircraft.
Aircraft autopilot systems use simple well proven sensors that transmit small amounts of data which is then processed by simple hardware and software that have been through years of rigorous testing, and there are multiple layers of redundancy. Software changes are rarely made. When they occur they are small, carefully executed, rigorously tested, and careful documented. All this has meant that software errors in aircraft autopilot systems have been extremely rare. (But have occurred, for instance Qantas Flight 72 from Singapore to Perth in 2008.)
Autonomous vehicles, in contrast, have “AI” systems that process up to gigabytes of sensor data each minute against deep learning artificial neural nets and “hand-crafted” code. The ANNs are effectively causal black boxes. Tesla software updates are done opaquely online to solve serious safety-related defects.
Tesla drivers and passengers have been referred to as guinea-pigs. Tesla calls it Autopilot software “Beta”. Yet this software is controlling vehicles on public roads, and Tesla recommends that customers to us its beta software on public roads.
The most startling difference between aircraft and vehicle autonomous control systems is the object recognition that’s crucial to vehicle safety but is lacking from aircraft systems because of the flight environment outside the aircraft.
Object recognition (in some sense of “recognition”) is essential in vehicle self-driving systems, and because of the great fragility and complexity of the recognition software, Tesla Autopilot is much more complex and less reliable than aircraft autopilot sysytems.
Elon Musk’s equating of Tesla Autopilot with aircraft autopilot is a dangerous deception.
Review of some of Mr Musk’s more concerning recent and less recent claims:
On Tesla Autopilot safety (Telegraph)
“The probability of having an accident is 50 per cent lower if you have Autopilot on,” said Musk, speaking at an energy conference in Oslo, Norway. “Even with our first version, it’s almost twice as good as a person.”
On the Tesla’s autonomous status (Road and Track):
“Even with the current sensor suite, we could make the car go fully autonomous…”
On Mr Musk’s unparalleled technical genius (Road and Track):
“I almost view it as a solved problem [how to build fully autonomous cars]. We know exactly what to do, and we’ll be there in a few years.”
These statements are dangerous. Now the corpses are piling up, regulatory bodies, consumer groups, and the responsible media are starting to challenge Musk and Tesla Inc. The worst part is that Mr Musk places a low value on human life.