In a demonstration of Mobileye’s autonomous vehicle technology, the company took their vehicle on a route through the middle of Jerusalem.
The autonomous vehicle features a camera-only system and maneuvers through busy intersections, around multi-lane roundabouts, and in other tricky situations that would challenge even the most adept human drivers.
The video shows the car from above thanks to a drone and the only time the demo had to pause was to recharge the drone’s battery.
The clip also features an internal camera in the car to show the human observer not touching the wheel, along with a digital representation in the top-left of what the computer vision system recognises, along with path planning.
Watching anyone drive for 40 minutes is not particularly fun, even a computer, so the footage is sped up to 2x speed.
What’s impressive about the demo is that shows an autonomous vehicle performing tasks that we’re not used to seeing cars be capable of.
These include capabilities that are not yet available even from Tesla:
- Turning corners at an intersection
- Navigating parked cars on the side of the street
- Merging into dense traffic on city streets
- Stopping for a pedestrian at a crossing
- Accelerating away at a green light
- Recognise the difference between a vehicle stopped in traffic, versus one that’s parked and making the appropriate lane changes.
- Cutting into traffic at T-intersection, forcing others to yield.
- Navigate around unusual objects (person with a forklift)
- Multi-lane roundabout
- Navigate double-parked vehicle
- Overtook into oncoming traffic (slow speeds)
Mobileye’s technology approach also differs dramatically from Tesla’s. While this demonstration was done using just their computer vision system, Mobileye wants to create an equally good version using lidar and radar. While that sounds expensive, compared to much cheaper cameras, they’re keen to have the redundancy confirm the computer vision system is working correctly.
In my mind I think they’ll end up moving away from that approach as they gain confidence in computer vision. What’s also a big differentiator is that Mobileye do still leverage High Definition maps of the city to assist the car to navigate. The issue with HD maps has always been their currency, with any intersection easily modified by councils etc and the map then becomes outdated and dangerous to rely on.
Mobileye is relying on crowdsourced information to help power this. Other vehicles using Mobileye ADAS technology, anonymously send data back to the cloud. Getting data enough cars, often enough to power the HD maps is the challenge, especially outside densely populated cities.
While this a really impressive demo from Mobileye, it is really important to appreciate that this is still a technology demo and not yet in production vehicles.
The system is also not perfect, at the 14 minute mark, where the car determines that a truck in front (no lights) is parked and the lights ahead have turned green, it begins to look at merging left to pass by the truck. As it goes to move, the other cars in the left lane had to avoid it. While there’s no audio feed from the car, I wouldn’t be surprised to cop a horn for this maneuver.
Shortly after it passes the truck, the car reaches a red light where the steering wheel is turning back and forward while the car waits for a green. This is really not the desired behaviour and would be quite unnerving to a driver/passenger.
At 19:55 seconds, the car navigates through another round about, impressive, but its decisions afterwards are not so positive. The car pauses as it negotiates a vehicle to its right. The pause is so extended that a car, likely frustrated by this point where a human would proceed, that it overtakes the Mobileye car.
Also, on a different note, having a drone flying above vehicles and houses like this, would be completely illegal in Australia.
You can watch the video from Mobileye below.