Uber's Self-driving AI Predicts the Trajectories of Pedestrians, Vehicles, and Cyclists (venturebeat.com) 44
In a preprint paper, Uber researchers describe MultiNet, a system that detects and predicts the motions of obstacles from autonomous vehicle lidar data. From a report: They say that unlike existing models, MultiNet reasons about the uncertainty of the behavior and movement of cars, pedestrians, and cyclists using a model that infers detections and predictions and then refines those to generate potential trajectories. Anticipating the future states of obstacles is a challenging task, but it's key to preventing accidents on the road. Within the context of a self-driving vehicle, a perception system has to capture a range of trajectories other actors might take rather than a single likely trajectory. For example, an opposing vehicle approaching an intersection might continue driving straight or turn in front of an autonomous vehicle; in order to ensure safety, the self-driving vehicle needs to reason about these possibilities and adjust its behavior accordingly.
MultiNet takes as input lidar sensor data and high-definition maps of streets and jointly learns obstacle trajectories and trajectory uncertainties. For vehicles (but not pedestrians or cyclists), it then refines these by discarding the first-stage trajectory predictions and taking the inferred center of objects and objects' headings before normalizing them and feeding them through an algorithm to make final future trajectory and uncertainty predictions. To test MultiNet's performance, the researchers trained the system for a day on ATG4D, a data set containing sensor readings from 5,500 scenarios collected by Uber's autonomous vehicles across cities in North America using a roof-mounted lidar sensor. They report that MultiNet outperformed several baselines by a significant margin on all three obstacle types (vehicles, pedestrians, and cyclists) in terms of prediction accuracies. Concretely, modeling uncertainty led to improvements of 9% to 13%, and it allowed for reasoning about the inherent noise of future traffic movement.
MultiNet takes as input lidar sensor data and high-definition maps of streets and jointly learns obstacle trajectories and trajectory uncertainties. For vehicles (but not pedestrians or cyclists), it then refines these by discarding the first-stage trajectory predictions and taking the inferred center of objects and objects' headings before normalizing them and feeding them through an algorithm to make final future trajectory and uncertainty predictions. To test MultiNet's performance, the researchers trained the system for a day on ATG4D, a data set containing sensor readings from 5,500 scenarios collected by Uber's autonomous vehicles across cities in North America using a roof-mounted lidar sensor. They report that MultiNet outperformed several baselines by a significant margin on all three obstacle types (vehicles, pedestrians, and cyclists) in terms of prediction accuracies. Concretely, modeling uncertainty led to improvements of 9% to 13%, and it allowed for reasoning about the inherent noise of future traffic movement.
Blah blah blah marketing (Score:4, Insightful)
Uber researchers said their invention was great, and better than other people's invention.
PR hot air.
Re:Blah blah blah marketing (Score:4, Interesting)
Inferring the intended direction of obstacles (and not just their current trajectories) is a fundamental aspect of a self-driving system, and is extensively used. So weird to see them talk about it like they just came up with this new thing.
You don't specifically train to "obvious signals" - for example, you don't write an algorithm to detect a flashing turn signal, or whatnot. You just train it with data involving vehicles and pedestrians, starting before some given action was taken and ranging through after it was taken, and let the neural net learn from any and all context clues in the scene. Maybe a car starting to hug the side of its lane, or slowing down, or the silhouette of the driver inside the vehicle turning their head is a clue that they're about to change lanes; whatever it is, you leave it up to the neural net to figure out what the best clues are.
Re: (Score:2)
Yep, I remember a couple of years ago, one of the companies (google I think) had produced a video with a sort of wireframe 3d model of everything the car detected and where it predicted they were going. And at least I have confidence that google isn't going to just disregard the output of their trajectory prediction because they think it will make the breaking more erratic.
Re: (Score:2)
Yep, I remember a couple of years ago, one of the companies (google I think) had produced a video with a sort of wireframe 3d model of everything the car detected and where it predicted they were going.
At least five years ago. Probably more.
Re: (Score:2)
The true innovation in Uber's research is obviously using this for actively targetting pedestrians and cyclists, an early prototype was demonstrated a couple of years back...
Re: (Score:2)
Re: (Score:3)
Don't other Self Driving AI do this?
The trick is to identify what type of object it found, then figure out their average speed of movement and assume a linear course.
The problem that happens, is people who do something unexpected. For me as a driver, this normally requires me to look at the face and body language of the possible accident. I see that biker looking lower down, he may be planning on stopping. The pedestrian is looking at a cell phone while walking, then they might go into traffic.
I can s
ESP (Score:2)
I sometimes look at the car in front and think "he's going to change lane soon". Maybe I delude myself that I have some supernatural power but it often seems to be true, something about the car's movement, a slight veer toward the lane then back again perhaps as maybe the driver checks a mirror. Then a little later comes the lane change.
Obviously there are no turn signals involved, this is still America, turn signals are just apologizing for changing lanes.
Maybe if I practice staring at goats long enough I
Re: (Score:3)
The use of turn signals may be construed as providing information to the enemy.
Maybe if I practice staring at goats long enough I will be able to force drivers to change lanes just with the power of my mind...
I've actually done something like this on occasion. I've got a sixth sense about assholes. And when I come upon someone trying to play roadblock, I flip on my turn signal as if to change lanes. They see this and quickly pull into the next lane first. Then I just gun it and continue straight ahead.
Re: (Score:2)
The use of turn signals may be construed as providing information to the enemy.
Maybe if I practice staring at goats long enough I will be able to force drivers to change lanes just with the power of my mind...
I've actually done something like this on occasion. I've got a sixth sense about assholes. And when I come upon someone trying to play roadblock, I flip on my turn signal as if to change lanes. They see this and quickly pull into the next lane first. Then I just gun it and continue straight ahead.
I can grin at the poetic justice of that. Still, I'd rather we all use turn-signals the way they were intended: to communicate intent. I have a bugaboo about drivers who don't use turn-signals. Using them deceptively could create a far worse problem.
Re: (Score:2)
Confirmation bias assuming everyone else is an asshole, driving too fast, "gunning it", ... I don't have full confidence in this assessment but you might want to check if *you're* the asshole.
Re: (Score:1)
driving too fast, "gunning it"
OK, boomer.
Re: ESP (Score:1)
I've got a sixth sense about assholes.
Smelling, touching and tasting them aren't enough?!
Poor choice of words, I'd think (Score:5, Funny)
"Uber's Self-driving AI Predicts the Trajectories of Pedestrians, Vehicles, and Cyclists"
Is that their trajectories after the Uber vehicle has struck them?
Re: (Score:2)
That's easy to code.
pedestrian.trajctory = (0,0,-6)
Re: (Score:2)
Re: (Score:1)
Re: (Score:2)
Cyclists + Pedestrians makes it reject the object till it's to late
Uber... doing our part to solve the overpopulation (Score:2)
âoeUber's Self-driving AI Predicts the Trajectories of Pedestrians, Vehicles, and Cyclistsâ
So it can figure out who to aim for, I guess?
Re: Uber... doing our part to solve the overpopula (Score:2)
const sanitized = userInput .replace(/[\u2018\u2019]/g, "'") .replace(/[\u201C\u201D]/g, '"');
One line of code. Thatâ(TM)s all I ask.
Alert, experienced drivers do this already (Score:2)
As you overtake a car, (or if you're filtering between them on a (motor)bike, are you looking at where the driver is looking? That's where she'll be going next, unless they're looking at their damn phone, in which case anything can happen. Can you see their front wheel? Is it turning? Where on the road is the car positioned? Is it drifting left and thus about to change lanes without checking mirrors or indicating? (Answer, yes, probably). Are we coming up to an exit and is the dick in the red Mustang
Re: (Score:2)
... and is the dick in the red Mustang that just blasted past about to smash on his brakes, cut across all lanes, and exit. Like he does every morning? ...
*lol* Next thing you know the AI transmits the license plate to other cars and soon all traffic halts, because AIs gotten scared and Mustang guy parts all traffic like Moses parted the Red Sea!
Sorry, my bad! Still learning!! (Score:2)
To let an AI make predictions is a contraction to the very idea of safety. Accidents happen often not because a situation was predicted, but because it wasn't predicted. The only smart AI here would be the one that doesn't try to predict what's going to happen next. Or what will an AI tell the driver when it predicted the traffic wrong? ...
Re: (Score:2)
Re: (Score:2)
To let an AI make predictions is a contraction to the very idea of safety. Accidents happen often not because a situation was predicted, but because it wasn't predicted. The only smart AI here would be the one that doesn't try to predict what's going to happen next. Or what will an AI tell the driver when it predicted the traffic wrong? ...
The AI could try to predict ALL the possible trajectories but it would be impossible to drive downtown where there are pedestrians if you are trying to avoid all possible accidents. At any given moment, a pedestrian could decide to jump out in front of traffic. Likewise with cars on a 2 lane highway. At any given moment, an oncoming car could decide to swerve into your lane. Any driver human or AI can only react to normal expected probabilities and only react to the unexpected when it starts to deviate
Re: (Score:2)
False. You can't stop and eliminate ALL risk - but you can certainly mitigate risks based on observed behaviour around you.
- start to merge (or at least be *ready* to merge - make sure your lane change is clear, etc) or slow (or at least release accelerator ) if something involving toddlers is happening, especially if they look uncontrolled
- same as above for some cyclists. Note, I generally assume more professional looking cyclists will do more often the right thing (by professional I mean look like race
Re: (Score:2)
No. You've got it exactly wrong. You expect everyone to be a proactive driver when you already know not everyone is.
For example, it's quite intelligent of us to drive in big SUVs and not on bikes, because we are safer in a big car than on a bike. Cyclists may not like it, but it is nevertheless a smart decision.
If you then start to swerve to avoid a collision with a bike then you have a chance to put yourself into danger. You might lose control over your car and so can an AI. It is generally safer to drive
Re: (Score:2)
Do you work for Uber's self-driving division?
Re:Sorry, my bad! Still learning!! (Score:4, Insightful)
To let an AI make predictions is a contraction to the very idea of safety. Accidents happen often not because a situation was predicted, but because it wasn't predicted. The only smart AI here would be the one that doesn't try to predict what's going to happen next. Or what will an AI tell the driver when it predicted the traffic wrong?
There's two parts to this, one is worst case performance where you have to assume any idiot can do anything physics permits. The other part is being a "normal" driver picking up on common hints and cues without being spooked by harmless events. If you practice defense driving - or just driving in general - you're very rarely at the edge of the control envelope where only a full emergency stop can prevent you from hitting the pedestrian crossing the road. You start braking down as the likelihood goes up, sometimes you catch it early and break smoothly other times you catch it late and break less smoothly. A human drives quite differently if you look like you're 95% likely to cross or 5% likely to cross. It doesn't mean you run down 1/20 pedestrians.
Re: (Score:2)
This is because (most) humans know something AIs don't: we know we have no control over the future, but only over the present.
Re: (Score:2)
As a driver, whether you are a human or a computer, and you are traveling faster than the car in front of you, without a prediction about future state, what action do you think would be taken?
Predictions are the very heart of successful navigation in the environment for animals and computers.
Re: (Score:2)
Obviously am I driving in the fast lane and am about to overtake the other car. You didn't predict this as a possible answer, did you?
Algorithms are not 'reasoning' (Score:5, Informative)
Also, guaranteed, it's 'predictive algorithm' is probably wrong more often than right because something that can't 'think' or 'reason' has no capability to understand human beings.
Re: (Score:2)
Wow! Uber invented vectors!!! (Score:1)
This is amaaaaaazing!!!!!
I am amazed!
Zero chance Uber will build and deploy an autonomous level 5 taxi before they run out of cash. Bye Uber but thanks for the cheap rides and the research work you did on the way down!
So, is this like a target aquistion system... (Score:1)
Dammit! (Score:2)
Do we need to go to actor school first, to be able to commit suicide by throwing us in front of an Uber?
Awesome! (Score:2)
Classic math problem (Score:2)
WTF? (Score:2)
I would say that