Driverless cars may need to drive more like humans
A hit-and-run incident that left a pedestrian gravely injured in San Francisco earlier this week is raising questions about whether autonomous vehicles (AVs) can handle the unexpected as well as, or better than, human drivers.
Driving the news: The incident involved both a human-driven car (which made the initial impact with the pedestrian) and a Cruise AV (which then also struck the victim).
Why it matters: Driverless cars need to do three things: See their environment, predict what's about to happen, and decide what to do.
- Prediction is the most challenging because AVs lack the instincts of a human driver.
The big picture: The promise that driverless cars will be safer than human drivers is almost the entire selling point of the technology — but as they roll out in a growing number of cities, a string of incidents in early markets like San Francisco and Austin has put the public on edge.
- This most recent incident is likely going to exacerbate that lack of trust, regardless of who is to blame.
- And it begs the question, would a human driver have behaved differently?
Details: Video footage viewed by Axios shows the pedestrian in the crosswalk being struck by a human-driven car, then bouncing off that car's windshield into the path of the robotaxi.
- In a statement, Cruise said its robotaxi "braked aggressively to minimize the impact" but was unable to stop before rolling over the woman and coming to a halt.
The intrigue: Police are still investigating what happened, but their star witness could wind up being the robotaxi itself.
- Like other AVs, Cruise's car is outfitted with multiple cameras and sensors, which recorded the incident.
- Cruise turned over the video footage to police and is cooperating with the investigation, a company spokesperson said.
Zoom in: In addition to camera footage, Cruise showed Axios a replay of what the AV's software recorded.
- The cars, represented by colored rectangles, were stopped at a red light. The human-driven car was in the left lane, the empty robotaxi in the right.
- When the light turned green, both cars drove straight ahead, through the intersection.
- The robotaxi was already tracking the pedestrian up ahead, represented by a triangle on the computer screen.
- The pedestrian stepped into the roadway from the right, crossing the path of the approaching robotaxi.
- She had cleared the right lane where the robotaxi was traveling, and was almost halfway across the intersection when she was struck by the human-driven car in the left lane. After the initial impact, she was flung back into the right lane, where she was run over by the robotaxi.
What they're saying: Experts say the robotaxi likely did as it was trained, slamming its brakes the instant it detected an obstacle — in this case, the pedestrian after she was struck by the other car.
Yes, but: What about the moments leading up to the accident?
- Even though its lane was clear, should the Cruise vehicle have been more cautious, knowing a pedestrian was crossing against the light on a busy street at night?
"When you see something really bad happening on the roadway, a reasonable driver would slow down as a precaution," said AV expert Philip Koopman, an associate professor at Carnegie Mellon University.
- "A reasonable driver doesn't say, 'That's not my problem. I'm just gonna keep going.'"
- "You need to predict the behavior of all the objects in the world you could come in contact with," added John Krafcik, former CEO of Waymo, a Cruise competitor.
The catch: Self-driving cars can only make decisions based on what they're programmed to do.
- AVs aren't coded "to anticipate a human projectile coming from the other way," said Jennifer Dukarski, a Michigan lawyer specializing in AVs.
What to watch: Dukarski wonders whether a new approach to autonomous driving from Tesla CEO Elon Musk could make AVs safer.
- Instead of being based on hundreds of thousands of lines of code with specific instructions for various foreseen scenarios, the latest version of Tesla's "Full Self-Driving" (FSD) technology is teaching itself how to drive by processing billions of frames of video of how humans actually drive.
- "It's like ChatGPT, but for cars," Tesla engineer Dhaval Shroff explained in Walter Isaacson's new Musk biography.
The bottom line: There's almost an infinite list of things that can happen on the road — which is why teaching AVs to prepare for the unexpected is so difficult.