Google has been successfully testing their fully-autonomous cars for a while, and now they’re ready to give them a big behavioral upgrade. So what are they changing?
According to Chris Urmson, Google’s man behind the plan for their self-driving vehicle efforts, the studies found the cars to be too cautious in real world environments. This behavioral change will see the cars drive more human-like by allowing the programs driving the car to cut corners, cross dividing lines to avoid road obstructions, and brake in a more fluid and natural way.
The Wall Street Journal pointed out a tendency for the vehicles to brake dramatically, and react to possible hazards in a seemingly-unnatural and awkward manner.
From their article:
During a recent test drive involving two Wall Street Journal reporters, the Google Lexus RX450h autonomous vehicle jabbed or tapped the brakes at seemingly odd times.
In one instance, the Lexus hit the brakes because it interpreted that a car approaching fast from behind might cut it off as it passed. The other car didn’t do that, creating a sense that the Google car slowed for no reason.
Another time, the car stopped at a busy T-intersection with limited views of cross traffic. It waited for about 30 seconds, began to make a left turn, but stopped in the middle of the intersection as a woman on the other side of the street walked near the edge of the sidewalk.
Google’s goal is to program the vehicles to ease into and out of braking, as many human motorists do. We humans tend to ease off of the brakes to show our intention to other drivers, rather than going from a dead stop to gunning it in a jerky and possibly dangerous manner. Rather than simply wait at four-way stops for the intersection to clear, the cars can creep forward to show their intention to proceed.
Google’s self-driving vehicles have been involved in 16 reported accidents, 12 of which were the car being rear-ended by other motorists. Google claims that all 12 incidents were when the car was stationary, and not a result of the awkward breaking technique that has been changed.
The cars can now adapt to the behavior of other motorist behaviors on the road in unique situations as well. The example given in the Wall Street Journal is a traffic light that is out. If the other cars on the road treat the intersection as a four-way stop, the Google car will adapt this methodology as well.
Another issue with the current programming of the vehicles is the strict guidelines the cars follow for turning. The software calculates a minimum distance from obstacles like oncoming traffic and roadside objects to determine a safe maneuver. The result is a awkward driving-line that is harder for other motorists to predict and react to.
The remedy is to allow the cars to “cheat” in the corners by taking the turning angles more directly and limiting the often too-wide turns the vehicles are currently programmed to make, as presented in this diagram from the WSJ article:
The vehicles have also adapted the ability to cross solid divider lines in the road - a capability that wasn’t initially programmed into the car to prevent it from straying from its lane.
The issue became apparent when the car’s lane was blocked by a parked car. The computer hadn’t been programmed to adapt to such a situation, so the Google car would just sit indefinitely in the road. Now it has a method of determination allowing it to cross into another lane to avoid obstacles like parked cars or other obstructions.
Google’s ultimate goal is to fully assimilate the vehicles into everyday driving situations and get them on the market. Ironically it seems the answer to successful autonomous vehicles isn’t to make them perfect, but rather more human.