Google teaching autonomous vehicles to share road safely with cyclists
11 July 2016
In the latest monthly report on its autonomous driving project, Google engineers outlined some of the steps they are taking to teach their self-driving cars to share the road safely with cyclists.
Google currently has 24 Lexus RX450h SUVs and 34 prototype vehicles on the road with autonomous driving capability. The Google autonomous fleet has so far racked up 1,725,911 miles driven in autonomous mode, with 1,158,921 miles driven manually. The fleet is averaging 15K to 17K autonomous miles per week, with testing locations in Mountain View, California; Kirkland, Washington; Phoenix, Arizona; and Austin, Texas.
The Google autonomous driving system recognizes cyclists as unique users of the road, whom the software treats conservatively. Among the examples cited:
When the sensors detect a parallel-parked car with an open door near a cyclist, the autonomous car is programmed to slow down or nudge over to give the rider enough space to move towards the center of the lane and avoid the door.
Google autonomous cars give cyclists ample buffer room when passing.
Google autonomous cars won’t squeeze by when cyclists take the center of the lane, even if there’s technically enough space.
The sensors can detect a cyclist’s hand signals as an indication of an intention to make a turn or shift over. The software is designed to remember previous signals from a rider so it can better anticipate a rider’s turn down the road.
Using machine learning, Google engineers have trained the software to recognize many different types of bikes—from multicolored frames, big wheels, bikes with car seats, tandem bikes, conference bikes, and unicycles.
|More than 100 riders pedal around autonomous car near Google’s campus. Click to enlarge.|
In the reports, Google lists any collisions—even minor. For June, these consisted of:
6 June 2016: A Google prototype autonomous vehicle (Google AV) in Austin sustained a small scrape to its front right fender after another vehicle approaching the Google AV from behind in an adjacent right turn-only lane crossed into the lane occupied by the Google AV and made slight contact with the side of the vehicle.
The other vehicle sustained a scrape on its left rear quarter panel. No injuries were reported
15 June 2016: Another Google AV in Austin was rear-ended. The Google AV was stopped at a red light for around one minute when the vehicle immediately behind the Google AV rolled forward and collided with the Google AV. The speed of the other vehicle at the time of the collision was approximately 3 mph. The Google AV sustained a minor scrape on its rear bumper. There was no visible damage to the other vehicle.
Google and Tesla should partner on this. Tesla has thousands of test vehicles on the road at any time vs. Google's 58 vehicles. By partnering Google gets thousand of test vehicles and Tesla would benefit from Google's more rapid self-driving capabilities.
This is the right thing to do, but competitive pressure may make this a non-starter. Hey Google and Tesla - think about it! Would be great for everybody!!
Posted by: Juan Valdez | 11 July 2016 at 09:01 PM
Juan I have had the exact same thoughts and like you I believe competition is preventing any corporation. No cooperation between the two may not be a bad thing after all because they are perusing very different ways toward the same goal and that may lead to faster development of the tech should one of the ways be a total flop. Google is making something that works nearly flawlessly in a geographically limited area that is mapped in high detail whereas Tesla is making a system that from the start works everywhere, and does not use highly detailed maps but that also is far from flawless. Google will expand their system by mapping more geographic areas and Tesla will keep eliminating the many situations that their AP system still cannot handle. Tesla is using maps but these maps are simple crowd-sourced maps generated from Tesla own vehicles and comparatively simple sensors whereas Google use a separate vehicle with highly sophisticated sensors like their high definition LIDAR to do some high definition mapping at first.
I think Tesla’s approach is the right one because Google will not be allowed by any authority to start a self-driving taxi service until they can document that it is safer than the average human driver by a significant margin like 10% safer. You need 5 billion miles and about 50 death accidents to prove that and google not even haing 2 million miles on the log is a real problem for Google.
How does Google go from where they are now with 2 million test miles and to the needed 5 billion test miles? And they need to do it by 2020 because that is when I expect Tesla to be ready with their driverless taxi service.
Musk has said that only one company keeps him up at nigh without saying which one. I say it is Google. The old auto industry seems aimless and paralyzed about driverless BEV services and Apple has nothing to show yet and therefore zero experience so far in this business.
Also read that Tesla’s next AP system (probably for summer 2017) will get 8 cameras, 1 radar and 12 ultrasonic sensors. Current AP has the same but only 1 camera.
Posted by: Account Deleted | 12 July 2016 at 12:44 AM
Tesla doesn't have the hardware on the cars that Google is heavily relying on - LIDAR
Posted by: GasperG | 12 July 2016 at 12:58 AM
The problem with cyclists is how to handle them on narrow two lane roads. In this case, the cyclist moving at say 15 mph is obstructing the flow of traffic at 40 mph. A motorist will squeeze past, or take a risk and overtake. A google car might stay behind for too long a time and cause tailbacks.
I am not anti-cyclist (as I cycle to work each day), but I can see problems on narrow, open roads.
(Roads with cycle-paths will obviously be OK).
Lets see how long it takes for a Tesla to kill / injure a cyclist. (I am not being smart here, Tesla is racking up far more real test miles and will encounter accidents more quickly (in terms of calendar time) than any other maker. The price for allowing this should be to force all companies to share their accident (and near miss?) data including video and all other sensor data from say 20 seconds before the accident to the impact or near miss.
Someone might need to define a multimedia interchange format to allow this... (Video, 3d video, radar, lidar, ultrasonics, all at different frame rates ...)
A nice little problem, but one that needs to be addressed so we can plug anyone's black box data into a reader, (and other companies algorithms).
Posted by: mahonj | 12 July 2016 at 01:40 AM
Tesla will have nearly about 300k vehicles with autopilot on the road by the end of 2017. Assume each do 10k miles per year in autopilot mode and 5k miles in manual driving per year. However, Tesla is logging all miles to analyse if the AP mode could have prevented accidents when not used and an accident occurs. So effectively Tesla will be logging accident statistics for 4.5 billion miles in 2018 (4,500,000,000=15,000*300,000) just from the cars that they sold before dec 31, 2017. Then you add the miles from the possibly 350k other cars they will sell in 2018 and Tesla will have the needed accident statistics to prove that their AP system is ready for large scale deployment as a driverless BEV taxi service.
I can’t see that Tesla can’t do this before 2020. However, the AP hardware system that will enable fully driverless BEVs will first deploy sometime in 2017 and there will not be enough of these cars on the road to make billions of miles of statistics until the end of 2018 so we need 2019 for data gathering as well so that full deployment of driverless taxi systems can begin in 2020.
@mahonji we need to ensure the privacy of people when publishing data. They have to be anonymous out of respect for the dead and the injured. So no video feeds just anonymous data. Also Tesla should not IMO be forced to give out their data until there are at least one more of their competitors that can also be forced to publish an equal amount of data. To ensure that will happen sooner rather than later I would support legislation that compelled all new cars to have autopilot by 2019 and also be compelled to log all accidents and publish the data on a monthly basis. If Tesla is the only one being forced to publish meaningful amounts of data because no other automaker including Google has got them it will evolve into a witch hunt where Tesla’s data is used to make accusations that are false and without scientific basis and where Tesla will not be able to compare their data with others in order to defend themselves.
Posted by: Account Deleted | 12 July 2016 at 02:26 AM
I gave it a little more thought. I think Google’s approach will flop if they try to get the old automakers to use their AP system because they will not. None of the old automakers are interested in being dependent on Google to deliver their AP system because it is too important for them and their survival to outsource that tech and they know it.
So the old automakers will go it alone by developing their own AP systems much like Tesla is doing. So far they are all behind Tesla as they lack a useful AP user interface (apart from Tesla’s UI it is all very confusing and too complicated for anyone to use on a daily basis) and none of the old automakers’ AP systems can be updated OTA. Nor do they collect accident statistics as Tesla’s AP system does and therefore they can’t develop their AP systems as fast as Tesla does. Currently the old automakers are not catching up to Tesla they are falling more behind and that will continue until OTA AP system updates and automatic statistics collection is implemented.
I expect Google at this time has realized that trying to sell their system to other automakers is mission impossible and therefore have made another business plan about making the system and the cars themselves. How that is going to happen I have no clue. They will need hundred of thousands of cars to prove that their system is working and currently they have less than 100. I see no other alternative for Google than to buy an existing automaker and then start outfitting their combustion cars with their AP system and also go full speed making BEV power trains. Volvo would be a good fit for Google but I do not think they are for sale at the moment.
Posted by: Account Deleted | 12 July 2016 at 03:37 AM
What does Tesla's stoplight mapping, detection and decision-making architecture look like? I was reading Google's first published paper on this topic in 2010; haven't seen or heard a thing from Tesla about traffic signal interaction.
Anticipation of driver reaction in other cars? The most recent Google car fender-bender was because of a mistaken "assumption" by the Google system (it guesses "other vehicle yields" and was wrong).
What is Tesla doing about all the interactions with other vehicles (fully- partially- or non-automated, big or small, cyclist or otherwise)?
Tesla fans are so obsessed with mapping that they don't seem to understand the totality of what Google is doing. Tesla's work is infantile by comparison, and their approach to tossing it out to drivers to "see how it goes" as a beta is not an acceptable development scheme for life-critical software.
Posted by: Herman | 12 July 2016 at 12:11 PM
"Nor do [Google] collect accident statistics as Tesla’s AP system does and therefore they can’t develop their AP systems as fast as Tesla does."
Nonsense. Google's data collection on driver take-over events, near-miss and accidents is not only extensive but in the state of California is reported in detail as an official submittal. Tesla's reporting? Zero.
Posted by: Herman | 12 July 2016 at 12:13 PM
Herman you are making up your own nonsense by “quoting” me wrongly either out of stupidity or because you are trying to manipulate. I expect you are just stupid so to make it crystal clear for you “they” refer to old auto industry not Google as is quite clear if you cared to read what I actually wrote.
Also Herman you don’t get that Google and Tesla have two opposite approaches to developing AP systems. Tesla makes an AP that works everywhere globally one feature at a time and Google makes an AP system that works with every feature accounted for one geographic place at a time. Tesla will do the traffic lights feature when they have the hardware and software that can do that feature on a global scale and they are not there yet as it is a really hard problem to solve on a global scale.
Google’s main problem is that they need at least 5 billion miles and 50 death accidents on the record before they are allowed to launch a driverless taxi service. So Google needs 300k data collecting vehicles to do 5 billion miles per year and they need to hurry because Tesla will get 300k AP data collecting vehicles by the end of 2017.
Posted by: Account Deleted | 12 July 2016 at 01:16 PM
How many miles per death is the average American getting? how many are these companies getting?
I think the Tesla to Google thing is apples to oranges really.
Google is trying to make a fully autonomous vehicle, tesla has advanced lane keeping and cruise control currently.
Several OEMs have been building/ trying to build autonomous vehicles since the 80s, I don't think there are any clear winners in this race to autonomy, but there are several approaching the finish line. (I believe most are waiting to cross it)
Depending how the Tesla autopilot death is handled... autonomy might be lost to lawyers in a courtroom/boardroom.
Henrik, I don't think 5 billion highway miles will make Tesla better than Google in the long run, again different strategies/ capabilities. Please don't make autopilot out to be something it is not.
Google doesn't have to do anything but sell what it has learned to the OEMs,car or no car, they are doing what they do best, data mining. The hardware to make autonomy possible is out now, its just implementation, and that is what Google and their trials, and other OEMs out there is doing.
Anyone can make a car follow the rules and the path of the road. Not everyone can account for people as of yet. Google and many others are trying to do that. Autonomous vehicles will launch in rapid succession right after the first public sale and about a years time to observe.
Litigation is probably the only thing holding many of these companies back. That and the laws aren't really agreement for autonomy just yet.
Posted by: CheeseEater88 | 12 July 2016 at 08:14 PM
CE88 in the US you get one death per 94 million miles on average and in the world it is one death per 60 million miles. The US has about 30,000 traffic deaths per year and the world has about 1.3 million traffic deaths per year. Tesla so far has only one death per 130 million miles on autopilot but it is only one death and you need about 50 deaths to be able to calculate an average and say with a high degree of certainty whether that average is representing a general picture of death on average or is just a coincidence. In order to get about 50 deaths you need to drive 5 billion miles.
So Google, Tesla or any other automaker will need to log 5 billion miles of detailed accident statistics before they can document in a scientifically correct manner that there auto-pilot systems are safer than humans behind the wheels. CE88 you should trust me on this one I have a PhD in data analysis it is what I do. So there is no way around it. Google or any other auto-company must get their auto-pilot system into minimum 300,000 vehicles that will start producing about 5 billion miles per year of accident statistics.
The tech race in the auto-industry is all about being first with fully autonomous vehicles. That is Google’s goal and that is Tesla’s goal but as I explained above they are approaching that goal differently. The Tesla AP is already much more than a sophisticated cruise control. It can change lanes, auto park and it can be summoned. An upgrade next month (AP version 8) will enable it to do off-ram exits on high-ways. It does operate both on highways and on ordinary roads but it is not yet capable of handling intersections and crossroads. These features will not arrive until after a hardware upgrade sometime in 2017 that will see Tesla apply 8 cameras instead of the one camera on the current Tesla AP system in addition to 12 ultrasonic and one radar sensor.
Posted by: Account Deleted | 12 July 2016 at 11:50 PM
I remember Google rejecting ULTRA PRT concept due to need constructing separate infrastructure for autonomous vehicles. Was it mistake? Still not sure. The designated light infrastructure for certain weight autos within town has economic advantages and could eliminate bottlenecks. Enginers know cost differnce between auto and railway bridge. The PRT bridges could hang even on house walls. In addition the concept would facilitate transition to autonomous street driving making hybrid systems from the start. At some road gaps even TESLA could be autonomous. I would like pointing out that ULTRA PRT is certified and implemented at small scale in London Heatrow airport. Making fully functional pure street autonomous driving still seems to me overambicious. IMO more rational would be step by step i.e. PRT way of development.
Posted by: Darius | 13 July 2016 at 12:24 AM
Collision Avoidance systems like precharging brakes, and applying them in certain cases will likely skew the numbers a bit. Again, statistics PhD or no, there are recognized differences in the levels of autonomy. From no autonomy, to full autonomy and every thing in between.
in 2018 my we will see auto braking standard on most cars sold in the US. This alone will save a tremendous amount of lives, being able to react when the driver is either unaware or cannot act fast enough.
Volvo has promised no more deaths in its future vehicles, while its a ridiculous claim, they might statistically get close.
Autonomous braking, is the first step.
Lane keeping/ autopilot type autonomy is close behind.
Full autonomy like Google and some other makers see is still a ways away, but everything they need to make the vehicle is here and now... its a infrastructure / programming problem. Governments would need to be involved, laws changed, exceptions made.
Crash avoidance systems/auto braking will be some of the easiest to implement, and have the largest pay off in reducing traffic deaths. Again, this is coming soon, to many makes.
Auto pilot like Tesla and others have for highway cruising add on to this. So its not really fair to look at them as separate.
The full on autonomous, uses everything at the car's disposal to choose the best course of action for the desired result. It is fully independent of the driver and their abilities, and can be the sole actor.
Test miles only on the freeway are not indicative of the full driving experience. Does tesla have access to all the data for every mile ever in auto pilot? does it accurately represent the average driver? (location, demographics, trips?)
More miles and more vehicles are helpful, but only if there is unlimited access to the data. Privacy and other rights (which may be waved) might stand in the way. Tesla and Google will gleam different data from their works. Some will overlap (especially on freeway travel) but they at this time are pursuing very different paths.
Google is looking to do something disruptive to the act of what we know as driving. Tesla is acting within the confines of that idea. Laws and perceptions don't need to change for an autopilot system that travels on the freeway. Autonomous vehicles will certainly be disruptive.
Most of the benefits will come from auto braking, the rest is mere convenience and revolution.
Posted by: CheeseEater88 | 13 July 2016 at 08:14 PM