The Flaw with Tesla Summon
… and other autonomous systems.
We like to believe that everything is in order. Indeed, in a world where visionary ideas and topics can quickly attract attention, making big claims has never been easier. With this recent trend, however, comes a troubling problem: execution.
The big self-driving car companies of today (Waymo, Tesla, and Uber, to name a few) have made large claims. A couple of years ago, big names in the Valley claimed that we’d have Level 4 autonomous cars by today. It’s just lane detection, sign classification, and cruise control, right?
Wrong. As engineers continue to work on autonomous driving, they are beginning to realize the true significance of the task. Even at MIT, engineers of robotics and autonomy are proposing that truly “solving” self-driving may not even be possible! Although there are many unpredictabilities in the realm of roadway driving, one of the biggest challenges are self-driving cars’ own creators: humans.
To be fair, artificial intelligence can learn a lot about a person, and quickly. Facial recognition is currently spreading throughout China. Anyone who has followed the current technology news can you tell you about AI beating humans at chess and Go.
Yet, predicting human behavior in a real life intersection is much more difficult than predicting a human’s next move in chess, even for fellow human beings. In general, people give off many signals, which can be split into two categories: direct communication (DC) and indirect communication (IC). DC, when a human explicitly says what they will do, is generally simple to predict for both humans and computers. The observer merely has trust that the pedestrian will do what they say. And if the pedestrian for some reason doesn’t do what they say they will, an accident could result. Even worse, IC, when the human gives off indirect signals/messages, is much harder to predict. In the context of predicting whether a pedestrian will cross, cues could be body rotation, head rotation, distance from the sidewalk, and whether they are on their phone.
Before reliable, Level 4 and 5 self-driving cars are released to the public, self-driving cars must prove themselves capable of fully interacting with unpredictable pedestrians. With research, investments, and patience, that may very well be possible. Or not.