Autonomous Car Progress

[ThomasD]

It will be really bad if this were to happen when there are thousands of these cars on the road

 

[KLam]

Pretty sure Elon said in an interview recently that driving comfort is currently being deprioritized, and they are focused on zero disengagements in urban areas first. So I don't think the issue is a lack of control system expertise.

At the same time, I don't buy that smoother control can't be improved in parallel with proper decision-making. Which implies that Tesla might be keeping the system jarring to make sure people stay alert while testing.

I hope so. But its just illogical for them to do that to generate more disengagements which may mask out some true problems.

 

[heltok]

Quick review of Volvo’s latest Lane Keep Assist, compared to Tesla it:
* Fails silently and surprisingly, feels like a terrible design choice
* Stops steering siliently when driving ≥140km/h. Set speed to 137km/h, it somehow went above 140km/h by itself, then steering is silently gone. Comes back when speed is reduced to 136km/h.
*Pingpongs a few decimeters in clearly marked lanes
*Really struggles with badly marked lanes
*Stops steering the moment blinker is turned on, takes back control after lane change.
*When lane marking disappears it stops steering without giving any warning

Tesla’s solution feels so much better, like Elon said, they are more than 5 years ahead of the competition.

 

[diplomat33]

Quick review of Volvo’s latest Lane Keep Assist, compared to Tesla it:
* Fails silently and surprisingly, feels like a terrible design choice
* Stops steering siliently when driving ≥140km/h. Set speed to 137km/h, it somehow went above 140km/h by itself, then steering is silently gone. Comes back when speed is reduced to 136km/h.
*Pingpongs a few decimeters in clearly marked lanes
*Really struggles with badly marked lanes
*Stops steering the moment blinker is turned on, takes back control after lane change.
*When lane marking disappears it stops steering without giving any warning

Tesla’s solution feels so much better, like Elon said, they are more than 5 years ahead of the competition.

Where is this review coming from? Is this your personal on-hands review? What is your source?

 

[heltok]

Where is this review coming from? Is this your personal on-hands review? What is your source?

It was me driving it for a week.

 

[diplomat33]

Waymo dialed up the confidence since the JJRicks videos, and probably has gone further since. There's an unprotected left at 3:00 where it looks a tiny bit twitchy while waiting but does not proceed until clear. It starts the turn while the last oncoming vehicle is still in the intersection, more aggressive than I recall. It makes a left in front of an oncoming car at 5:20. It wasn't dangerous, but rattled the couple. I don't remember JJ's cars being nearly that aggressive.

You've probably already seen this video but it has 2 more examples of Waymo being more confident/assertive in unprotected left turns.

 

[Dan D.]

You've probably already seen this video but it has 2 more examples of Waymo being more confident/assertive in unprotected left turns.

I hadn't watched that one before. The turn at 0:45 is just barging through a line of cars. I'd call that a taxi move in Vancouver, and if the other car hadn't turned left I'd say it was a d**kh**d turn. I couldn't tell if the car knew the other car was turning left when it went in front or if it was going to make its way through anyway.

I guess aggressive moves may be useful to make forward progress sometimes, but I'd keep my eye on this type of behaviour.

 

[diplomat33]

I hadn't watched that one before. The turn at 0:45 is just barging through a line of cars. I'd call that a taxi move in Vancouver, and if the other car hadn't turned left I'd say it was a d**kh**d turn. I couldn't tell if the car knew the other car was turning left when it went in front or if it was going to make its way through anyway.

To the human eye, I agree that it can look a d**k move. But I think the Waymo likely knew that the white Tesla was going to turn left. I say that because of all the cameras, radar and lidar on the Waymo and all the behavior prediction that Waymo has done. Waymo can predict paths of hundreds of objects up to 8 seconds into the future. So the Waymo Driver likely had a very precise position, velocity and prediction for the white Tesla and knew it was going to turn.

With the power of machine learning and large enough data on human driving, neural nets could be very good at predicting behavior. Also, autonomous cars with cameras, lidar and radar have the potential for superhuman perception. And computers can do trillions of operations per second. So autonomous cars have the potential to be far better than human drivers, not just with vision, but also at anticipating behavior. The only kink could be when humans decide to do something unpredictable. That is why AVs should have defensive driving skills.

I guess aggressive moves may be useful to make forward progress sometimes, but I'd keep my eye on this type of behaviour.

If autonomous cars are too timid, they could get stuck with indecisiveness and get "bullied" by human drivers. We've all driven in places where sometimes you need to be assertive or you won't get anywhere. So AVs do need to learn to be somewhat aggressive in some situations. Of course, if AVs are too assertive or aggressive, they could cause accidents. So AVs definitely need to find the right balance between caution and assertiveness.

 

[MrHappy]

Waymo must have precogs that can predict driver inentions before they happen.

 

[diplomat33]

Waymo must have precogs that can predict driver inentions before they happen.

It's not precog. It's just really good perception and really good machine learning. If you have really good perception, not just on the exact position and velocity of another car, but also things like micro movements of the wheels and contextual information like is the other vehicle in a turning lane, you can collect lots of data and train deep neural networks to learn to predict what other road agents might do. Obviously, no prediction will be 100% perfect all the time but you can get predictions that are very good.

Waymo has done lots of machine learning on behavior prediction. In fact, they've presented 6 research papers this year already on behavior prediction. For example, here is the abstract for their research paper on StopNet, a NN that can predict motion of hundreds of road agents:

We introduce a motion forecasting (behavior prediction) method that meets the latency requirements for autonomous driving in dense urban environments without sacrificing accuracy. A whole-scene sparse input representation allows StopNet to scale to predicting trajectories for hundreds of road agents with reliable latency. In addition to predicting trajectories, our scene encoder lends itself to predicting whole-scene probabilistic occupancy grids, a complementary output representation suitable for busy urban environments. Occupancy grids allow the AV to reason collectively about the behavior of groups of agents without processing their individual trajectories. We demonstrate the effectiveness of our sparse input representation and our model in terms of computation and accuracy over three datasets. We further show that co-training consistent trajectory and occupancy predictions improves upon state-of-the-art performance under standard metrics.

Research — StopNet: Scalable Trajectory and Occupancy Prediction for Urban Autonomous Driving — Waymo

We introduce a motion forecasting (behavior prediction) method that meets the latency requirements for autonomous driving in dense urban environments without sacrificing accuracy. A whole-scene sparse input representation allows StopNet to scale to predicting trajectories for hundreds of road...

waymo.com

 

[Daniel in SD]

Waymo must have precogs that can predict driver inentions before they happen.

The Tesla driver is also a precog!
Seriously though I think the Waymo saw the Tesla slow down and angle slightly to the left. It's also quite possible that they signaled and it doesn't show up video due to frame rate.

 

[diplomat33]

Mobileye launches "EyeQ Kit", a software development kit that will help automakers build their own safety and driver assist systems:

What’s new: Mobileye, an Intel company, has launched the EyeQ Kit™ – its first software development kit (SDK) for the EyeQ® system-on-chip that powers driver-assistance and future autonomous technologies for automakers worldwide. Built to leverage the powerful and highly power-efficient architecture of the upcoming EyeQ®6 High and EyeQ®Ultra processors, EyeQ Kit allows automakers to utilize Mobileye's proven core technology, while deploying their own differentiated code and human-machine interface tools on the EyeQ platform.

How it Works: Through EyeQ hardware and software, automakers have access to a broad set of Mobileye solutions, including computer vision, REM™ crowdsourced mapping, and RSS-based driving policy. Using EyeQ Kit, automakers can further leverage the power of Mobileye’s system-on-chip to augment the advanced driver functions with a look and feel that is unique to their brands. And as the visual demands for interaction and communication between drivers and vehicles grows more complex, EyeQ Kit gives automakers a new path to tailor critical information flows. EyeQ Kit will help support features such as surround visualization, automated lane-keeping, and road-sign recognition through more advanced augmented reality displays.

EyeQ Kit has already been deployed with one major global automaker for future vehicle programs.

Mobileye Launches EyeQ Kit: New SDK for Advanced Safety and...

Toolkit builds on two decades of Mobileye safety and driver-assistance technology to provide new flexibility for future software-defined vehicles.

www.intel.com

 

[diplomat33]

Some more details on Zoox's progress. They finished testing the Zoox driverless vehicle that has no steering wheel or pedals on open private roads and will begin testing on public roads next year in SF:

Levinson said the vehicle has been driven “with no one inside, no chase vehicle, and no emergency stop, all on open, private roads with non-Zoox agents, including pedestrians, cyclists, cars and trucks,” opening the way to the next phase of development and testing. Safety requirements have been met, he said, and the vehicle can now operate in an unstructured environment at a safety level above that of humans.

The next steps are now to test the vehicle, which has been redesigned from the ground up to seat four people, has no driver’s seat, and can drive in both directions, on public roads. This is to begin next year in San Francisco, where adapted test vehicles are already on the road.

Starting from these vehicle tests, a successively larger fleet will then be on the road with passengers, which will then be transferred to a robotaxi service.

Zoox Reaches Important Milestone

Zoox, a San Francisco Bay Area-based startup that is developing self-driving technology as well as its own vehicle and was purchased by Amazon two years ago for about $1 billion, has reached a majo…

thelastdriverlicenseholder.com

 

[AlanSubie4Life]

Some more details on Zoox's progress. They finished testing the Zoox driverless vehicle that has no steering wheel or pedals on open private roads and will begin testing on public roads next year in SF:

Zoox Reaches Important Milestone

Zoox, a San Francisco Bay Area-based startup that is developing self-driving technology as well as its own vehicle and was purchased by Amazon two years ago for about $1 billion, has reached a majo…

thelastdriverlicenseholder.com

I wonder how they’re going to manually unstick giant Zoox traffic jams? It’ll be a real zoo of Zoox.

I guess they probably have some sort of portable console control, maybe. They must of course.

 

[diplomat33]

I wonder how they’re going to manually unstick giant Zoox traffic jams? It’ll be a real zoo of Zoox.

Why would there be a giant Zoox traffic jams? Cruise had traffic jams with their vehicles but that does not necessarily mean Zoox will have the same issue.

I guess they probably have some sort of portable console control, maybe. They must of course.

Zoox uses teleguidance. So a person monitors the vehicle and provides guidance to the vehicle, but does not directly control the vehicle. But perhaps for this driverless vehicle, since it has no steering wheel or pedals, they will add some sort of console remote control for emergencies if teleguidance does not work.

 

[AlanSubie4Life]

Why would there be a giant Zoox traffic jams?

Seems more likely than not. Just sort of the way things go with automated systems. But yeah it is certainly possible they’ll find a way to avoid it.

 

[Dewg]

As an aside to the current discussion - I've often wondered this: I know it's illegal, but there are signal jammers out there. If someone engaged a signal jammer and disrupted all cellular and RF signals to any of the autonomous vehicles, I wonder what would happen.

Would they continue, using onboard systems to navigate to their destination? Would they enter an emergency mode and pull over safely and stop?

I was thinking about a security conference last year that showed people hacking various cars and controlling functions. I couldn't imagine the damage someone could do if they got into the remote control functions of a Waymo or Cruise.

 

[diplomat33]

As an aside to the current discussion - I've often wondered this: I know it's illegal, but there are signal jammers out there. If someone engaged a signal jammer and disrupted all cellular and RF signals to any of the autonomous vehicles, I wonder what would happen.

Would they continue, using onboard systems to navigate to their destination? Would they enter an emergency mode and pull over safely and stop?

I was thinking about a security conference last year that showed people hacking various cars and controlling functions. I couldn't imagine the damage someone could do if they got into the remote control functions of a Waymo or Cruise.

Waymo and Cruise can do full autonomous driving with no cell signal or internet connection because everything needed for autonomous driving is on board the car. So they would not be affected by signal jammers. Waymo and Cruise also have firewalls to protect from hacking. Also, Waymo and Cruise don't have any remote control functions. So there is no remote control functions to hack.

 

[Dewg]

Waymo and Cruise can do full autonomous driving with no cell signal or internet connection because everything needed for autonomous driving is on board the car. So they would not be affected by signal jammers. Waymo and Cruise also have firewalls to protect from hacking. Also, Waymo and Cruise don't have any remote control functions. So there is no remote control functions to hack.

Excellent! I always thought there was some form of remote assistance they had - such as if the car ran into a situation, like an accident that blocked the entire road, where a safety driver could take over remotely and navigate the car around the obstacle. I may have that totally wrong.

 

[diplomat33]

Excellent! I always thought there was some form of remote assistance they had - such as if the car ran into a situation, like an accident that blocked the entire road, where a safety driver could take over remotely and navigate the car around the obstacle. I may have that totally wrong.

As explained before, the cars do have remote assistance but they do not remote control the car, they just provide guidance to the car. Remote assistance can drop waypoints on the map to suggest a path the car should take. Remote assistance can also label objects to help the autonomous driving understand the object's intent. The car stays in autonomous mode and treats the guidance like perception data to try to figure out what to do. The remote assistance do not actually control the car.

This video explains how it works with some examples: