{"id":139,"date":"2026-08-29T08:28:56","date_gmt":"2026-08-29T08:28:56","guid":{"rendered":"https:\/\/www.graveiensai.com\/blog\/?p=139"},"modified":"2026-08-29T08:28:56","modified_gmt":"2026-08-29T08:28:56","slug":"ai-in-self-driving-cars","status":"publish","type":"post","link":"https:\/\/www.graveiensai.com\/blog\/ai-in-self-driving-cars\/","title":{"rendered":"What is AI in Self-Driving Cars? How Autonomous Vehicles Work in 2026"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\">AI in self-dri\u00adving cars is the stack of machine learn\u00ading sys\u00adtems that let a vehi\u00adcle sense its sur\u00adround\u00adings, pre\u00addict what road users will do next, and con\u00adtrol steer\u00ading, brak\u00ading, and accel\u00ader\u00ada\u00adtion with\u00adout a human dri\u00adver. In plain terms, arti\u00adfi\u00adcial intel\u00adli\u00adgence is the \u201cbrain\u201d that turns raw sen\u00adsor sig\u00adnals into safe dri\u00adving deci\u00adsions many times per sec\u00adond. This guide explains what that brain is made of, how do self dri\u00adving cars work end to end, the six lev\u00adels of dri\u00adving automa\u00adtion, the sen\u00adsor and soft\u00adware choic\u00ades that sep\u00ada\u00adrate the lead\u00aders from the rest, and the train\u00ading data that qui\u00adet\u00adly decides whether any of it is safe enough to trust.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">You will get a plain-lan\u00adguage def\u00adi\u00adn\u00adi\u00adtion, a com\u00adpar\u00adi\u00adson of the main tech\u00adni\u00adcal approach\u00ades, an orig\u00adi\u00adnal readi\u00adness mod\u00adel you can reuse, a prac\u00adti\u00adcal check\u00adlist, real safe\u00adty num\u00adbers from pri\u00adma\u00adry sources, and answers to the ques\u00adtions peo\u00adple actu\u00adal\u00adly ask. Whether your inter\u00adest is full auton\u00ado\u00admy or the auto\u00admo\u00adtive automa\u00adtion already ship\u00adping in today\u2019s cars, the same prin\u00adci\u00adples apply.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>At a glance<\/strong><\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th><strong>Ques\u00adtion<\/strong><\/th><th><strong>Direct answer<\/strong><\/th><\/tr><\/thead><tbody><tr><td>What is AI in self-dri\u00adving cars?<\/td><td>The per\u00adcep\u00adtion, pre\u00addic\u00adtion, plan\u00adning, and con\u00adtrol soft\u00adware (most\u00adly deep learn\u00ading) that replaces the human driver\u2019s eyes, judg\u00adment, and hands.<\/td><\/tr><tr><td>How do self-dri\u00adving cars work?<\/td><td>A four-stage loop: sense the world with sen\u00adsors, per\u00adceive and clas\u00adsi\u00adfy objects, pre\u00addict their motion, then plan and exe\u00adcute a dri\u00adving path.<\/td><\/tr><tr><td>What are the lev\u00adels?<\/td><td>SAE International\u2019s J3016 stan\u00addard defines six lev\u00adels of dri\u00adving automa\u00adtion, from Lev\u00adel 0 (no automa\u00adtion) to Lev\u00adel 5 (full automa\u00adtion any\u00adwhere).<\/td><\/tr><tr><td>Is it safe yet?<\/td><td>In its oper\u00adat\u00ading areas, Way\u00admo reports 94% few\u00ader seri\u00adous-injury crash\u00ades than human dri\u00advers across 220.6 mil\u00adlion rid\u00ader-only miles (through March 2026).<\/td><\/tr><tr><td>What decides suc\u00adcess?<\/td><td>Sen\u00adsor choice and, above all, the qual\u00adi\u00adty and cov\u00ader\u00adage of the labeled train\u00ading data behind the per\u00adcep\u00adtion sys\u00adtem.<\/td><\/tr><tr><td>Who leads in 2026?<\/td><td>Way\u00admo (US rob\u00ado\u00adt\u00adaxis), Baidu Apol\u00adlo Go (Chi\u00adna), and Tesla\u2019s vision-first approach are the most cit\u00aded pro\u00adgrams; GM wound down its Cruise rob\u00ado\u00adt\u00adaxi effort in late 2024.<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>Table of contents<\/strong><\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li>What is AI in self-dri\u00adving cars?<\/li>\n\n\n\n<li>How do self-dri\u00adving cars work?<\/li>\n\n\n\n<li>The six lev\u00adels of dri\u00adving automa\u00adtion<\/li>\n\n\n\n<li>The sen\u00adsor suite: how an auto\u00admat\u00aded vehi\u00adcle sees<\/li>\n\n\n\n<li>LiDAR-first vs vision-first: a nuanced com\u00adpar\u00adi\u00adson<\/li>\n\n\n\n<li>Auto\u00admo\u00adtive automa\u00adtion and the data behind it<\/li>\n\n\n\n<li>The Graveiens AV Data Readi\u00adness Matrix<\/li>\n\n\n\n<li>How to eval\u00adu\u00adate an autonomous dri\u00adving data pro\u00adgram<\/li>\n\n\n\n<li>A worked exam\u00adple: the occlud\u00aded pedes\u00adtri\u00adan<\/li>\n\n\n\n<li>What the safe\u00adty data shows about AI in self-dri\u00adving cars<\/li>\n\n\n\n<li>Fre\u00adquent\u00adly asked ques\u00adtions<\/li>\n\n\n\n<li>About the authors<\/li>\n\n\n\n<li>Con\u00adclu\u00adsion<\/li>\n\n\n\n<li>Sources<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>What is AI in self-driving cars?<\/strong><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>AI in self-dri\u00adving cars refers to the fam\u00adi\u00adly of machine learn\u00ading mod\u00adels that per\u00adceive the envi\u00adron\u00adment, antic\u00adi\u00adpate oth\u00ader road users, and decide how the vehi\u00adcle should move, all in real time.<\/strong> It is not one algo\u00adrithm but a pipeline of spe\u00adcial\u00adized sys\u00adtems work\u00ading togeth\u00ader.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">A use\u00adful men\u00adtal mod\u00adel is to think of the human tasks a dri\u00adver per\u00adforms and the AI com\u00adpo\u00adnent that replaces each one. Your eyes become cam\u00aderas, radar, and LiDAR. Your visu\u00adal under\u00adstand\u00ading becomes a per\u00adcep\u00adtion net\u00adwork that detects and clas\u00adsi\u00adfies objects. Your antic\u00adi\u00adpa\u00adtion of what oth\u00ader dri\u00advers might do becomes a pre\u00addic\u00adtion mod\u00adel. Your deci\u00adsion to slow, stop, or steer becomes a plan\u00adning and con\u00adtrol sys\u00adtem. The role of arti\u00adfi\u00adcial intel\u00adli\u00adgence in self-dri\u00adving cars is to run this loop reli\u00adably in con\u00addi\u00adtions no engi\u00adneer can ful\u00adly script in advance, from a plas\u00adtic bag blow\u00ading across a high\u00adway to a child step\u00adping out between parked cars.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">This mat\u00adters because dri\u00adving is an open-end\u00aded prob\u00adlem. Tra\u00addi\u00adtion\u00adal rule-based soft\u00adware strug\u00adgles with the long tail of rare events. Mod\u00adern autonomous sys\u00adtems lean on deep learn\u00ading trained on enor\u00admous vol\u00adumes of real and sim\u00adu\u00adlat\u00aded dri\u00adving data, which is why the qual\u00adi\u00adty of that data, cov\u00adered lat\u00ader in this guide, is wide\u00adly regard\u00aded as one of the most impor\u00adtant fac\u00adtors in whether a pro\u00adgram suc\u00adceeds. This shift from hand-writ\u00adten rules to learned behav\u00adior is what mod\u00adern auto\u00admo\u00adtive automa\u00adtion real\u00adly means, and it is the rea\u00adson data has become the cen\u00adtral bat\u00adtle\u00adground for auto\u00admat\u00aded vehi\u00adcles.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>How do self-driving cars work?<\/strong><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Self-dri\u00adving cars work through a con\u00adtin\u00adu\u00adous four-stage loop: sens\u00ading, per\u00adcep\u00adtion, pre\u00addic\u00adtion, and plan\u00adning with con\u00adtrol.<\/strong> Engi\u00adneers often short\u00aden this to \u201csense, think, act.\u201d Under\u00adstand\u00ading how do self dri\u00adving cars work at this lev\u00adel is the fastest way to under\u00adstand where AI adds val\u00adue and where it can fail.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The loop runs many times each sec\u00adond and looks like this.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">1. <strong>Sense. <\/strong>Cam\u00aderas, radar, LiDAR, ultra\u00adson\u00adic sen\u00adsors, GPS, and iner\u00adtial units cap\u00adture a raw, over\u00adlap\u00adping pic\u00adture of the world. NVIDIA describes this mul\u00adti-sen\u00adsor input as the foun\u00adda\u00adtion that lets the car \u201csee\u201d in con\u00addi\u00adtions no sin\u00adgle sen\u00adsor han\u00addles alone.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">2. <strong>Per\u00adceive. <\/strong>Neur\u00adal net\u00adworks turn those raw sig\u00adnals into a struc\u00adtured scene: this pix\u00adel clus\u00adter is a cyclist, that box is a bus, this region is dri\u00advable road. Tech\u00adniques such as object detec\u00adtion and pix\u00adel-lev\u00adel seman\u00adtic seg\u00admen\u00adta\u00adtion do the heavy lift\u00ading here.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">3. <strong>Pre\u00addict. <\/strong>The sys\u00adtem fore\u00adcasts where each detect\u00aded agent is like\u00adly to move over the next few sec\u00adonds. A pedes\u00adtri\u00adan fac\u00ading the curb behaves dif\u00adfer\u00adent\u00adly from one already step\u00adping into the road.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">4. <strong>Plan and act. <\/strong>A plan\u00adning mod\u00adule choos\u00ades a safe, com\u00adfort\u00adable tra\u00adjec\u00adto\u00adry, and con\u00adtrol sys\u00adtems trans\u00adlate it into steer\u00ading, throt\u00adtle, and brak\u00ading com\u00admands.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The Way\u00admo Dri\u00adver Hand\u00adbook frames per\u00adcep\u00adtion as the stage where an auto\u00admat\u00aded vehi\u00adcle builds a real-time under\u00adstand\u00ading of every\u00adthing around it, then con\u00adtin\u00adu\u00adal\u00adly updates that under\u00adstand\u00ading as the scene changes. Every lat\u00ader deci\u00adsion is only as good as this per\u00adceived pic\u00adture, which is why so much engi\u00adneer\u00ading effort, and so much labeled data, con\u00adcen\u00adtrates on the first two stages.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>The six levels of driving automation<\/strong><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>SAE International\u2019s J3016 stan\u00addard defines six lev\u00adels of dri\u00adving automa\u00adtion, num\u00adbered 0 to 5, describ\u00ading how much of the dri\u00adving task the sys\u00adtem han\u00addles and when a human must be ready to take over.<\/strong> These lev\u00adels are the shared lan\u00adguage reg\u00adu\u00adla\u00adtors, car\u00admak\u00aders, and the press use, so it is worth know\u00ading them pre\u00adcise\u00adly.<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th><strong>Lev\u00adel<\/strong><\/th><th><strong>Name<\/strong><\/th><th><strong>Who dri\u00adves<\/strong><\/th><th><strong>Exam\u00adple<\/strong><\/th><\/tr><\/thead><tbody><tr><td>0<\/td><td>No automa\u00adtion<\/td><td>Human, always<\/td><td>Basic warn\u00adings, emer\u00adgency brak\u00ading that only assists<\/td><\/tr><tr><td>1<\/td><td>Dri\u00adver assis\u00adtance<\/td><td>Human, with one aid<\/td><td>Adap\u00adtive cruise con\u00adtrol or lane keep\u00ading, not both<\/td><\/tr><tr><td>2<\/td><td>Par\u00adtial automa\u00adtion<\/td><td>Human super\u00advis\u00ades<\/td><td>Com\u00adbined steer\u00ading and speed assist; hands and eyes still required<\/td><\/tr><tr><td>3<\/td><td>Con\u00addi\u00adtion\u00adal automa\u00adtion<\/td><td>Sys\u00adtem, in set con\u00addi\u00adtions<\/td><td>Car dri\u00adves itself in defined sit\u00adu\u00ada\u00adtions but may ask the human to take over<\/td><\/tr><tr><td>4<\/td><td>High automa\u00adtion<\/td><td>Sys\u00adtem, with\u00adin its domain<\/td><td>Rob\u00ado\u00adt\u00adaxi that needs no human inside its mapped ser\u00advice area<\/td><\/tr><tr><td>5<\/td><td>Full automa\u00adtion<\/td><td>Sys\u00adtem, every\u00adwhere<\/td><td>No steer\u00ading wheel required, any road, any con\u00addi\u00adtion<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">A key nuance the SAE guid\u00adance stress\u00ades: Lev\u00adels 0 to 2 are dri\u00adver sup\u00adport fea\u00adtures, where a per\u00adson is always dri\u00adving even when their feet are off the ped\u00adals. Lev\u00adels 3 to 5 are auto\u00admat\u00aded dri\u00adving fea\u00adtures, where the sys\u00adtem is dri\u00adving when engaged. Most cars you can buy today sit at Lev\u00adel 2. The rob\u00ado\u00adt\u00adaxis oper\u00adat\u00ading in sev\u00ader\u00adal cities are Lev\u00adel 4, mean\u00ading a ful\u00adly auto\u00admat\u00aded vehi\u00adcle with\u00adin a spe\u00adcif\u00adic, mapped oper\u00adat\u00ading area rather than every\u00adwhere. No pro\u00adduc\u00adtion vehi\u00adcle has cred\u00adi\u00adbly reached Lev\u00adel 5.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>The sensor suite: how an automated vehicle sees<\/strong><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>An auto\u00admat\u00aded vehi\u00adcle per\u00adceives the world by fus\u00ading sev\u00ader\u00adal com\u00adple\u00admen\u00adtary sen\u00adsors, because no sin\u00adgle sen\u00adsor is reli\u00adable in all con\u00addi\u00adtions.<\/strong> This prin\u00adci\u00adple, called sen\u00adsor fusion, is cen\u00adtral to safe auton\u00ado\u00admy. The sen\u00adsor lay\u00ader is the eyes of the car, and get\u00adting it right is the first require\u00adment for a depend\u00adable auto\u00admat\u00aded vehi\u00adcle. The main sen\u00adsors each have dis\u00adtinct strengths and blind spots.<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th><strong>Sen\u00adsor<\/strong><\/th><th><strong>Strength<\/strong><\/th><th><strong>Weak\u00adness<\/strong><\/th><th><strong>Typ\u00adi\u00adcal role<\/strong><\/th><\/tr><\/thead><tbody><tr><td>Cam\u00adera<\/td><td>Rich col\u00ador and tex\u00adture, reads signs and lights, low cost<\/td><td>Poor depth, strug\u00adgles in glare, fog, and dark\u00adness<\/td><td>Clas\u00adsi\u00adfi\u00adca\u00adtion, traf\u00adfic lights, lane lines<\/td><\/tr><tr><td>Radar<\/td><td>Works in rain, fog, and dark; mea\u00adsures speed direct\u00adly<\/td><td>Low res\u00ado\u00adlu\u00adtion, coarse shape detail<\/td><td>Dis\u00adtance and clos\u00ading speed, adap\u00adtive cruise<\/td><\/tr><tr><td>LiDAR<\/td><td>Pre\u00adcise 3D depth and shape, day or night<\/td><td>High\u00ader cost, degrad\u00aded in heavy pre\u00adcip\u00adi\u00adta\u00adtion<\/td><td>3D map\u00adping, object size and posi\u00adtion<\/td><\/tr><tr><td>Ultra\u00adson\u00adic<\/td><td>Accu\u00adrate at very short range<\/td><td>Use\u00adless at dis\u00adtance<\/td><td>Park\u00ading, low-speed maneu\u00advers<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">Because each sen\u00adsor cov\u00aders another\u2019s weak\u00adness, most Lev\u00adel 4 pro\u00adgrams com\u00adbine all of them and merge their out\u00adputs into one con\u00adsis\u00adtent 3D mod\u00adel of the scene. Build\u00ading and label\u00ading that fused, mul\u00adti-sen\u00adsor view is a spe\u00adcial\u00adized dis\u00adci\u00adpline, and it is where 3D point cloud and LiDAR anno\u00adta\u00adtion work becomes essen\u00adtial. For teams build\u00ading per\u00adcep\u00adtion at this lev\u00adel, the accu\u00adra\u00adcy of fused sen\u00adsor labels is a make-or-break input, which is why ded\u00adi\u00adcat\u00aded <a href=\"https:\/\/www.graveiensai.com\/sensor-fusion-lidar\">sen\u00adsor fusion and LiDAR anno\u00adta\u00adtion<\/a> is treat\u00aded as core infra\u00adstruc\u00adture rather than an after\u00adthought. Our guide to <a href=\"https:\/\/www.graveiensai.com\/blog\/semantic-segmentation\/\">seman\u00adtic seg\u00admen\u00adta\u00adtion<\/a> goes deep\u00ader on how pix\u00adel-lev\u00adel label\u00ading sup\u00adports this stage. <em>(Also read.)<\/em><\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>LiDAR-first vs vision-first: a nuanced comparison<\/strong><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>There is no sin\u00adgle \u201cbest\u201d sen\u00adsor strat\u00ade\u00adgy: LiDAR-first and vision-first approach\u00ades each win in dif\u00adfer\u00adent con\u00addi\u00adtions, and many teams now blend them.<\/strong> Fram\u00ading this as one side being sim\u00adply cor\u00adrect mis\u00adreads the engi\u00adneer\u00ading real\u00adi\u00adty.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The LiDAR-first camp, which includes Way\u00admo and most rob\u00ado\u00adt\u00adaxi oper\u00ada\u00adtors, argues that pre\u00adcise 3D depth from LiDAR pro\u00advides a safe\u00adty mar\u00adgin that pure vision can\u00adnot yet match, espe\u00adcial\u00adly for rare, high-con\u00adse\u00adquence events. The vision-first camp, most asso\u00adci\u00adat\u00aded with Tes\u00adla, argues that cam\u00aderas plus pow\u00ader\u00adful neur\u00adal net\u00adworks can learn depth and con\u00adtext the way humans do, at a frac\u00adtion of the hard\u00adware cost, and that scale of data mat\u00adters more than exot\u00adic sen\u00adsors.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The hon\u00adest assess\u00adment is that each approach has gen\u00aduine trade-offs.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>LiDAR-first <\/strong>tends to win where absolute reli\u00ada\u00adbil\u00adi\u00adty in poor vis\u00adi\u00adbil\u00adi\u00adty and unusu\u00adal objects mat\u00adters most, and where per-vehi\u00adcle cost is sec\u00adondary to safe\u00adty head\u00adroom. Rob\u00ado\u00adt\u00adaxis fit this pro\u00adfile.<\/li>\n\n\n\n<li><strong>Vision-first <\/strong>tends to win where cost, scal\u00ada\u00adbil\u00adi\u00adty, and fleet-wide data col\u00adlec\u00adtion mat\u00adter most, and where the prod\u00aduct can improve grad\u00adu\u00adal\u00adly under human super\u00advi\u00adsion. Con\u00adsumer dri\u00adver-assis\u00adtance fits this pro\u00adfile.<\/li>\n\n\n\n<li><strong>Hybrid <\/strong>designs, increas\u00ading\u00adly com\u00admon, use cam\u00aderas for rich seman\u00adtic under\u00adstand\u00ading and radar or LiDAR for depend\u00adable depth, aim\u00ading to cap\u00adture the best of both.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">The prac\u00adti\u00adcal take\u00adaway is that the sen\u00adsor debate is down\u00adstream of a deep\u00ader ques\u00adtion: whichev\u00ader sen\u00adsors you choose, the mod\u00adels still have to be trained on data that rep\u00adre\u00adsents the messy real world. That is the con\u00adstant across every approach.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>Automotive automation and the data behind it<\/strong><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Auto\u00admo\u00adtive automa\u00adtion suc\u00adceeds or fails on train\u00ading data, not just algo\u00adrithms, because a per\u00adcep\u00adtion mod\u00adel can only rec\u00adog\u00adnize what it has been taught to see.<\/strong> This is the part of the sto\u00adry that gets the least atten\u00adtion and deserves the most.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">A mod\u00adern auto\u00admat\u00aded vehi\u00adcle can gen\u00ader\u00adate on the order of ter\u00adabytes of sen\u00adsor data per day of dri\u00adving. Turn\u00ading that raw stream into some\u00adthing a mod\u00adel can learn from requires dis\u00adci\u00adplined <a href=\"https:\/\/www.graveiensai.com\/data-collection\">data col\u00adlec\u00adtion<\/a>, pre\u00adcise <a href=\"https:\/\/www.graveiensai.com\/data-annotation\">data anno\u00adta\u00adtion<\/a>, and rig\u00ador\u00adous <a href=\"https:\/\/www.graveiensai.com\/data-validation\">val\u00adi\u00adda\u00adtion<\/a>. Each frame may need bound\u00ading box\u00ades, lane mark\u00adings, dri\u00advable-area masks, and 3D cuboids around vehi\u00adcles and pedes\u00adtri\u00adans, often across fused cam\u00adera and LiDAR views. This label\u00ading work is the foun\u00adda\u00adtion of the <a href=\"https:\/\/www.graveiensai.com\/computer-vision\">com\u00adput\u00ader vision<\/a> sys\u00adtems that let the car per\u00adceive its sur\u00adround\u00adings.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The hard\u00adest and most valu\u00adable data is the long tail: the rare, ambigu\u00adous, safe\u00adty-crit\u00adi\u00adcal moments that rarely appear in ordi\u00adnary dri\u00adving logs. A con\u00adstruc\u00adtion work\u00ader direct\u00ading traf\u00adfic by hand, an over\u00adturned couch on the free\u00adway, a scoot\u00ader run\u00adning a red light at dusk. Pro\u00adgrams that sys\u00adtem\u00adat\u00adi\u00adcal\u00adly mine, label, and sim\u00adu\u00adlate these edge cas\u00ades build a durable advan\u00adtage in <a href=\"https:\/\/www.graveiensai.com\/adas\">ADAS and autonomous<\/a> per\u00adcep\u00adtion, because those are exact\u00adly the sce\u00adnar\u00adios where a poor\u00adly trained mod\u00adel behaves unpre\u00addictably.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Human judg\u00adment stays in the loop through\u00adout. Skilled anno\u00adta\u00adtors resolve the cas\u00ades mod\u00adels find con\u00adfus\u00ading, review\u00aders catch label\u00ading errors before they poi\u00adson a dataset, and domain experts define what \u201ccor\u00adrect\u201d even means for a nov\u00adel scene. Our work on <a href=\"https:\/\/www.graveiensai.com\/blog\/physical-ai-robotics-training-data\/\">phys\u00adi\u00adcal AI and robot\u00adics train\u00ading data<\/a> explores how this same human-in-the-loop dis\u00adci\u00adpline extends from vehi\u00adcles to robots. <em>(Also read.) <\/em>When remote human oper\u00ada\u00adtors need to assist a stuck vehi\u00adcle, that fall\u00adback depends on its own care\u00adful\u00adly designed work\u00adflows, a top\u00adic we cov\u00ader in our piece on <a href=\"https:\/\/www.graveiensai.com\/blog\/teleoperation\/\">tele\u00adop\u00ader\u00ada\u00adtion<\/a>. <em>(Also read.)<\/em><\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>The Graveiens AV Data Readiness Matrix<\/strong><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Use this matrix to judge whether an autonomous dri\u00adving data pro\u00adgram is actu\u00adal\u00adly ready to sup\u00adport safe deploy\u00adment, rather than mere\u00adly pro\u00adduc\u00ading labels.<\/strong> We built this mod\u00adel from repeat\u00aded pat\u00adterns in per\u00adcep\u00adtion data work, because a program\u2019s readi\u00adness is easy to over\u00adstate and hard to mea\u00adsure. It scores five pil\u00adlars across three matu\u00adri\u00adty stages. Find your hon\u00adest posi\u00adtion in each row.<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th><strong>Data pil\u00adlar<\/strong><\/th><th><strong>Nascent<\/strong><\/th><th><strong>Devel\u00adop\u00ading<\/strong><\/th><th><strong>Deploy\u00adment-ready<\/strong><\/th><\/tr><\/thead><tbody><tr><td>Sce\u00adnario cov\u00ader\u00adage<\/td><td>Most\u00adly clear-weath\u00ader, com\u00admon roads<\/td><td>Some weath\u00ader and night data<\/td><td>Sys\u00adtem\u00adat\u00adic edge-case and long-tail min\u00ading<\/td><\/tr><tr><td>Anno\u00adta\u00adtion pre\u00adci\u00adsion<\/td><td>2D box\u00ades, incon\u00adsis\u00adtent class\u00ades<\/td><td>Con\u00adsis\u00adtent 2D plus basic 3D<\/td><td>Fused 2D and 3D cuboids, tight tol\u00ader\u00adances, agreed tax\u00adon\u00ado\u00admy<\/td><\/tr><tr><td>Sen\u00adsor fusion labels<\/td><td>Cam\u00adera only<\/td><td>Cam\u00adera plus one oth\u00ader sen\u00adsor<\/td><td>Time-synced cam\u00adera, LiDAR, radar labeled togeth\u00ader<\/td><\/tr><tr><td>Qual\u00adi\u00adty assur\u00adance<\/td><td>Sin\u00adgle-pass label\u00ading<\/td><td>Spot-check review<\/td><td>Mul\u00adti-stage QA with mea\u00adsured inter-anno\u00adta\u00adtor agree\u00adment<\/td><\/tr><tr><td>Feed\u00adback loop<\/td><td>No struc\u00adtured error min\u00ading<\/td><td>Occa\u00adsion\u00adal mod\u00adel-dri\u00adven rela\u00adbel\u00ading<\/td><td>Con\u00adtin\u00adu\u00adous active learn\u00ading that tar\u00adgets mod\u00adel fail\u00adures<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">The rule of thumb: a pro\u00adgram is only as strong as its weak\u00adest row. A team with beau\u00adti\u00adful 3D labels but no edge-case cov\u00ader\u00adage will still be sur\u00adprised on the road. Mov\u00ading every pil\u00adlar to \u201cdeploy\u00adment-ready\u201d is the real work of auto\u00admo\u00adtive automa\u00adtion, and it is rarely fin\u00adished, because the long tail nev\u00ader ful\u00adly ends. Pro\u00adgrams that treat auto\u00admo\u00adtive automa\u00adtion as a data prob\u00adlem, not just a soft\u00adware prob\u00adlem, tend to age bet\u00adter.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>How to evaluate an autonomous driving data program<\/strong><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>To judge an AV data pipeline, test it against the fail\u00adure modes that actu\u00adal\u00adly cause on-road inci\u00addents, not against its best-case demo.<\/strong> The reli\u00ada\u00adbil\u00adi\u00adty of an auto\u00admat\u00aded vehi\u00adcle is decid\u00aded far more by this pipeline than by any sin\u00adgle mod\u00adel choice, and the same holds for low\u00ader-lev\u00adel auto\u00admo\u00adtive automa\u00adtion fea\u00adtures. Use this check\u00adlist when you build or buy per\u00adcep\u00adtion data.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">1. <strong>Define the oper\u00ada\u00adtional design domain first. <\/strong>Know exact\u00adly where and when the vehi\u00adcle is meant to dri\u00adve before you col\u00adlect a sin\u00adgle frame.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">2. <strong>Audit sce\u00adnario cov\u00ader\u00adage. <\/strong>Ask what frac\u00adtion of the dataset is night, rain, fog, glare, and unusu\u00adal objects, not just clear-day high\u00adway.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">3. <strong>Insist on a writ\u00adten label tax\u00adon\u00ado\u00admy. <\/strong>Ambi\u00adgu\u00adi\u00adty in class def\u00adi\u00adn\u00adi\u00adtions is a lead\u00ading source of silent errors.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">4. <strong>Mea\u00adsure inter-anno\u00adta\u00adtor agree\u00adment. <\/strong>If two skilled anno\u00adta\u00adtors dis\u00adagree often, the guide\u00adlines, not the peo\u00adple, need fix\u00ading.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">5. <strong>Require mul\u00adti-stage qual\u00adi\u00adty assur\u00adance. <\/strong>One review\u00ader catch\u00ading another\u2019s mis\u00adtakes should be the norm, not the excep\u00adtion.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">6. <strong>Check sen\u00adsor syn\u00adchro\u00adniza\u00adtion. <\/strong>Fused labels are only use\u00adful if cam\u00adera and LiDAR frames are time-aligned to the mil\u00adlisec\u00adond.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">7. <strong>Close the loop with active learn\u00ading. <\/strong>The pipeline should keep sur\u00adfac\u00ading the exact frames where the cur\u00adrent mod\u00adel fails.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">8. <strong>Val\u00adi\u00addate on held-out edge cas\u00ades. <\/strong>Reserve rare, hard scenes the mod\u00adel has nev\u00ader seen to test whether it tru\u00adly gen\u00ader\u00adal\u00adizes.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Run any pro\u00adgram against these eight points and its real matu\u00adri\u00adty, in the lan\u00adguage of the readi\u00adness matrix above, becomes obvi\u00adous quick\u00adly.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>A worked example: the occluded pedestrian<\/strong><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Here is how bet\u00adter data changes an out\u00adcome: a pedes\u00adtri\u00adan part\u00adly hid\u00adden behind a parked truck at dusk is the kind of edge case that sep\u00ada\u00adrates a frag\u00adile mod\u00adel from a robust one.<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong><em>Before. <\/em><\/strong>A per\u00adcep\u00adtion mod\u00adel trained most\u00adly on clear-day footage sees only the pedestrian\u2019s legs below the truck. With no sim\u00adi\u00adlar exam\u00adples in its train\u00ading set, it labels the region low-con\u00adfi\u00addence \u201cunknown\u201d and the plan\u00adner treats the space as dri\u00advable. The safe\u00adty mar\u00adgin is thin.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong><em>After. <\/em><\/strong>The team mines thou\u00adsands of par\u00adtial-occlu\u00adsion scenes, labels them with con\u00adsis\u00adtent 3D cuboids across fused cam\u00adera and LiDAR, and adds match\u00ading sim\u00adu\u00adlat\u00aded vari\u00ada\u00adtions at dif\u00adfer\u00adent light lev\u00adels. Retrained, the mod\u00adel now rec\u00adog\u00adnizes \u201cpar\u00adtial\u00adly occlud\u00aded pedes\u00adtri\u00adan, like\u00adly to emerge,\u201d rais\u00ades its con\u00adfi\u00addence, and the plan\u00adner slows and widens its gap before the per\u00adson steps out.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Noth\u00ading about the algo\u00adrithm changed in this exam\u00adple. The dif\u00adfer\u00adence is entire\u00adly in the cov\u00ader\u00adage and pre\u00adci\u00adsion of the data, which is the recur\u00adring les\u00adson of real autonomous dri\u00adving work.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>What the safety data shows about AI in self-driving cars<\/strong><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>The best cur\u00adrent evi\u00addence sug\u00adgests mature Lev\u00adel 4 sys\u00adtems can be safer than human dri\u00advers with\u00adin their oper\u00adat\u00ading areas, though the data is still geo\u00adgraph\u00adi\u00adcal\u00adly nar\u00adrow.<\/strong> Attri\u00adbu\u00adtion mat\u00adters here, so these fig\u00adures come from pri\u00adma\u00adry sources.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Accord\u00ading to Waymo\u2019s pub\u00adlished safe\u00adty impact data, across 220.6 mil\u00adlion rid\u00ader-only miles through March 2026 in its oper\u00adat\u00ading cities, its vehi\u00adcles were involved in 94% few\u00ader seri\u00adous-injury-or-worse crash\u00ades and 82% few\u00ader any-injury-report\u00aded crash\u00ades than the human bench\u00admark for the same areas. Way\u00admo also reports 93% few\u00ader pedes\u00adtri\u00adan-injury crash\u00ades and 84% few\u00ader cyclist-injury crash\u00ades. These com\u00adpar\u00adisons are drawn from police-report\u00aded crash records and are lim\u00adit\u00aded to the spe\u00adcif\u00adic cities and con\u00addi\u00adtions where Way\u00admo cur\u00adrent\u00adly dri\u00adves.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">For the human base\u00adline, the fig\u00adure often quot\u00aded is from the US Nation\u00adal High\u00adway Traf\u00adfic Safe\u00adty Administration\u2019s Nation\u00adal Motor Vehi\u00adcle Crash Cau\u00adsa\u00adtion Sur\u00advey, which found the crit\u00adi\u00adcal rea\u00adson for a crash was assigned to the dri\u00adver in an esti\u00admat\u00aded 94% of cas\u00ades. NHTSA itself cau\u00adtions that this \u201ccrit\u00adi\u00adcal rea\u00adson\u201d is not the same as the cause of the crash, so the sta\u00adtis\u00adtic sup\u00adports the case for automa\u00adtion with\u00adout prov\u00ading that automa\u00adtion would pre\u00advent 94% of crash\u00ades. Respon\u00adsi\u00adble report\u00ading keeps that dis\u00adtinc\u00adtion.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The mar\u00adket is bet\u00adting heav\u00adi\u00adly on this tra\u00adjec\u00adto\u00adry. Prece\u00addence Research esti\u00admates the autonomous vehi\u00adcle mar\u00adket at USD 273.75 bil\u00adlion in 2025 and projects it could reach rough\u00adly USD 5.4 tril\u00adlion by 2035 at about a 34.84% com\u00adpound annu\u00adal growth rate, though such long-range pro\u00adjec\u00adtions vary wide\u00adly between research firms and should be read as direc\u00adtion\u00adal. Deploy\u00adment is uneven: Way\u00admo has expand\u00aded to more US cities in 2026 and stat\u00aded a goal of reach\u00ading one mil\u00adlion paid rides per week, Baidu\u2019s Apol\u00adlo Go runs one of the largest rob\u00ado\u00adt\u00adaxi fleets in Chi\u00adna, and Gen\u00ader\u00adal Motors wound down its Cruise rob\u00ado\u00adt\u00adaxi pro\u00adgram in late 2024, a reminder that the path to scaled auton\u00ado\u00admy is nei\u00adther smooth nor guar\u00adan\u00adteed.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>Frequently asked questions<\/strong><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>What is AI in self-dri\u00adving cars in sim\u00adple terms?<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">It is the soft\u00adware brain that replaces a human dri\u00adver. AI in self-dri\u00adving cars sens\u00ades the road with cam\u00aderas and oth\u00ader sen\u00adsors, fig\u00adures out what every\u00adthing is and what it will do next, and then steers, accel\u00ader\u00adates, and brakes accord\u00ading\u00adly, with\u00adout a per\u00adson con\u00adtrol\u00adling the vehi\u00adcle.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>How do self-dri\u00adving cars work step by step?<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">They run a repeat\u00ading loop: sen\u00adsors cap\u00adture the scene, per\u00adcep\u00adtion mod\u00adels iden\u00adti\u00adfy objects, pre\u00addic\u00adtion mod\u00adels fore\u00adcast move\u00adment, and a plan\u00adner choos\u00ades and exe\u00adcutes a safe path. This \u201csense, think, act\u201d cycle repeats many times per sec\u00adond, and it is the sim\u00adplest cor\u00adrect answer to how do self dri\u00adving cars work.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Who invent\u00aded self-dri\u00adving car tech\u00adnol\u00ado\u00adgy?<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">There is no sin\u00adgle inven\u00adtor. Ear\u00adly mile\u00adstones include Carnegie Mel\u00adlon and Mer\u00adcedes-Benz research in the 1980s and 1990s and the DARPA Grand Chal\u00adlenges of the mid-2000s, which spurred much of the mod\u00adern indus\u00adtry. Way\u00admo grew out of Google\u2019s self-dri\u00adving project that began in 2009.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>What are the lev\u00adels of self-dri\u00adving cars?<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">SAE International\u2019s J3016 stan\u00addard defines six lev\u00adels, from Lev\u00adel 0 (no automa\u00adtion) to Lev\u00adel 5 (full automa\u00adtion every\u00adwhere). Most new cars offer Lev\u00adel 2 dri\u00adver sup\u00adport; today\u2019s rob\u00ado\u00adt\u00adaxis are Lev\u00adel 4, mean\u00ading full automa\u00adtion only with\u00adin a defined area.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Is an auto\u00admat\u00aded vehi\u00adcle safer than a human dri\u00adver?<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">With\u00adin its lim\u00adit\u00aded oper\u00adat\u00ading areas, Way\u00admo reports far few\u00ader injury crash\u00ades than human dri\u00advers, includ\u00ading 94% few\u00ader seri\u00adous-injury crash\u00ades across 220.6 mil\u00adlion rid\u00ader-only miles. The evi\u00addence is encour\u00adag\u00ading but still geo\u00adgraph\u00adi\u00adcal\u00adly nar\u00adrow, so broad claims should be made care\u00adful\u00adly.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Do self-dri\u00adving cars use LiDAR or cam\u00aderas?<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">It depends on the com\u00adpa\u00adny. Most rob\u00ado\u00adt\u00adaxi oper\u00ada\u00adtors fuse LiDAR, radar, and cam\u00aderas for redun\u00addan\u00adcy, while Tes\u00adla favors a vision-first approach built main\u00adly on cam\u00aderas. Each strat\u00ade\u00adgy has real trade-offs in cost, reli\u00ada\u00adbil\u00adi\u00adty, and scal\u00ada\u00adbil\u00adi\u00adty.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Why is train\u00ading data so impor\u00adtant for self-dri\u00adving cars?<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Because a per\u00adcep\u00adtion mod\u00adel can only rec\u00adog\u00adnize what it has been taught to see. Broad, pre\u00adcise, well-val\u00adi\u00addat\u00aded train\u00ading data, espe\u00adcial\u00adly of rare edge cas\u00ades, is wide\u00adly regard\u00aded as one of the most deci\u00adsive fac\u00adtors in whether AI in self-dri\u00adving cars is safe.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>What is the dif\u00adfer\u00adence between ADAS and full auto\u00admo\u00adtive automa\u00adtion?<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">ADAS, or advanced dri\u00adver-assis\u00adtance sys\u00adtems, cov\u00aders the Lev\u00adel 1 and Lev\u00adel 2 fea\u00adtures such as adap\u00adtive cruise con\u00adtrol and lane keep\u00ading, where a human is always respon\u00adsi\u00adble. Full auto\u00admo\u00adtive automa\u00adtion refers to Lev\u00adel 4 and Lev\u00adel 5, where the sys\u00adtem dri\u00adves itself and no human super\u00advi\u00adsion is required inside its oper\u00adat\u00ading domain. Most cars today offer ADAS; a self-dri\u00adving auto\u00admat\u00aded vehi\u00adcle you can ride with no dri\u00adver exists only as a Lev\u00adel 4 rob\u00ado\u00adt\u00adaxi in select cities.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Is full self-dri\u00adving avail\u00adable to buy in 2026?<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">No pro\u00adduc\u00adtion vehi\u00adcle offers true Lev\u00adel 5 automa\u00adtion. Con\u00adsumers can buy Lev\u00adel 2 dri\u00adver assis\u00adtance, and Lev\u00adel 4 rob\u00ado\u00adt\u00adaxis oper\u00adate as a ser\u00advice in select cities rather than as a car you own and dri\u00adve any\u00adwhere.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>About the authors<\/strong><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">This guide was pro\u00adduced by the Graveiens AI Edi\u00adto\u00adr\u00adi\u00adal Team, which spe\u00adcial\u00adizes in the data that pow\u00aders per\u00adcep\u00adtion sys\u00adtems for autonomous dri\u00adving and robot\u00adics, and reviewed by [Senior Review\u00ader name and title], a data lead with [X]+ years in AI train\u00ading data and ADAS anno\u00adta\u00adtion. Graveiens AI is a human-in-the-loop data ser\u00advices com\u00adpa\u00adny deliv\u00ader\u00ading mul\u00adti\u00adlin\u00adgual data col\u00adlec\u00adtion, anno\u00adta\u00adtion, and val\u00adi\u00adda\u00adtion across auto\u00admo\u00adtive, health\u00adcare, finance, and geospa\u00adtial domains, with qual\u00adi\u00adty process\u00ades aligned to rec\u00adog\u00adnized stan\u00addards [con\u00adfirm cer\u00adti\u00adfi\u00adca\u00adtion, for exam\u00adple ISO 9001:2017]. Learn more on our <a href=\"https:\/\/www.graveiensai.com\/about-us\">about page<\/a>.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>Conclusion<\/strong><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The role of AI in self-dri\u00adving cars is to do reli\u00adably what a good human dri\u00adver does intu\u00aditive\u00adly: see the road, antic\u00adi\u00adpate oth\u00ader peo\u00adple, and act safe\u00adly, thou\u00adsands of times per trip. As this guide has shown, how do self dri\u00adving cars work comes down to a sense-think-act loop, the SAE lev\u00adels set expec\u00adta\u00adtions for what an auto\u00admat\u00aded vehi\u00adcle can and can\u00adnot do, and the sen\u00adsor debate is real but sec\u00adondary to one deep\u00ader truth. Every approach to auto\u00admo\u00adtive automa\u00adtion rests on the qual\u00adi\u00adty, breadth, and pre\u00adci\u00adsion of its train\u00ading data.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">That is where the hard\u00adest, most durable advan\u00adtage is built, and it is exact\u00adly the work Graveiens AI does for per\u00adcep\u00adtion teams: edge-case data col\u00adlec\u00adtion, fused 2D and 3D anno\u00adta\u00adtion, and rig\u00ador\u00adous val\u00adi\u00adda\u00adtion, with skilled humans in the loop at every stage. If you are build\u00ading or scal\u00ading an autonomous dri\u00adving pro\u00adgram, <a href=\"https:\/\/www.graveiensai.com\/contact-us\">talk to our team about a data pilot<\/a>.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>Sources<\/strong><\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Way\u00admo, Safe\u00adty Impact (rid\u00ader-only miles and crash com\u00adpar\u00adisons): <a href=\"https:\/\/waymo.com\/safety\/impact\/\" target=\"_blank\" rel=\"noopener\">https:\/\/waymo.com\/safety\/impact\/<\/a><\/li>\n\n\n\n<li>SAE Inter\u00adna\u00adtion\u00adal, J3016 Lev\u00adels of Dri\u00adving Automa\u00adtion: <a href=\"https:\/\/www.sae.org\/standards\/content\/j3016_202104\/\" target=\"_blank\" rel=\"noopener\">https:\/\/www.sae.org\/standards\/content\/j3016_202104\/<\/a><\/li>\n\n\n\n<li>US NHTSA, Crit\u00adi\u00adcal Rea\u00adsons for Crash\u00ades (DOT HS 812 506): <a href=\"https:\/\/crashstats.nhtsa.dot.gov\/Api\/Public\/ViewPublication\/812506\" target=\"_blank\" rel=\"noopener\">https:\/\/crashstats.nhtsa.dot.gov\/Api\/Public\/ViewPublication\/812506<\/a><\/li>\n\n\n\n<li>NVIDIA, How Does a Self-Dri\u00adving Car See: <a href=\"https:\/\/blogs.nvidia.com\/blog\/how-does-a-self-driving-car-see\/\" target=\"_blank\" rel=\"noopener\">https:\/\/blogs.nvidia.com\/blog\/how-does-a-self-driving-car-see\/<\/a><\/li>\n\n\n\n<li>Way\u00admo Dri\u00adver Hand\u00adbook, Per\u00adcep\u00adtion: <a href=\"https:\/\/waymo.com\/blog\/2021\/10\/the-waymo-driver-handbook-perception\/\" target=\"_blank\" rel=\"noopener\">https:\/\/waymo.com\/blog\/2021\/10\/the-waymo-driver-handbook-perception\/<\/a><\/li>\n\n\n\n<li>Prece\u00addence Research, Autonomous Vehi\u00adcle Mar\u00adket: <a href=\"https:\/\/www.precedenceresearch.com\/autonomous-vehicle-market\" target=\"_blank\" rel=\"noopener\">https:\/\/www.precedenceresearch.com\/autonomous-vehicle-market<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>AI in self-dri\u00adv\u00ading cars is the stack of machine learn\u00ading sys\u00adtems that let a vehi\u00adcle sense its sur\u00adround\u00adings, pre\u00addict what road users will do next, and con\u00adtrol steer\u00ading,\u2026<\/p>\n","protected":false},"author":1,"featured_media":140,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wp_typography_post_enhancements_disabled":false,"footnotes":""},"categories":[1],"tags":[],"class_list":["post-139","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-blog"],"_links":{"self":[{"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/posts\/139","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/comments?post=139"}],"version-history":[{"count":1,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/posts\/139\/revisions"}],"predecessor-version":[{"id":141,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/posts\/139\/revisions\/141"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/media\/140"}],"wp:attachment":[{"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/media?parent=139"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/categories?post=139"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.graveiensai.com\/blog\/wp-json\/wp\/v2\/tags?post=139"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}