drive.ai 的自動駕駛技術

這家公司的共同創辦人兼CEO Sameep Tandon告訴我們，目前他們的車隊已經具備L4級別（部分狀況下的完全自動駕駛）的自動駕駛水平！這是很特別的一種方式！

https://www.cnet.com/roadshow/news/drive-ai-brings-emotional-intelligence-to-self-driving-cars/

Self-driving car technology is being developed by most major automakers, automotive equipment suppliers and start-up companies such as Drive.ai. The intent is to eliminate the 95 percent of fatal car accidents every year that are attributable to human error.

Carol Reiley, President and co-founder of Drive.ai, said the 20 or so people on staff came from Stanford's Artificial Intelligence Lab, and decided to put their PhDs on hold to solve the problems self-driving cars will face. The company focuses on deep learning, sometimes referred to as machine learning, a relatively new area of computer science. With deep learning, rather than responding to a set of specific instructions, computers can extrapolate from prior experiences.

Using deep learning programming, a car's computer can accurately identify what its sensors perceive, then decide how to react to each situation. However, pedestrians in an urban environment can by very unpredictable in their movements, or, as Reiley puts it "people are very dynamic in their decision-making."

https://www.bnext.com.tw/article/43656/standford-autopilot-car-drive.ai
在自動駕駛領域，基本可以分成兩個流派：一個是採用經典機器人方向，是基於規則的（rule-base） 的。工程師會為每個場景都寫好固定的代碼，來告訴機器人應該怎麼去做。這樣的結果是，如果新的場景出現、又沒有對應代碼的話，那麼機器很可能就不知道怎麼應對。這就嚴重限制了它的可拓展性。

舉一個例子，Waymo的自動駕駛汽車，在從總部山景城擴展到奧斯汀的時候，僅僅因為山景城的紅綠燈是豎向的，而奧斯汀的則是橫向，就沒有辦法順利識別紅綠燈，而不得不讓程序員重新去寫程式「教」它。

另外一個現在更受歡迎、包括Drive.ai也選擇的方向，是基於深度學習技術。深度學習可以模擬大腦識別機制，對於非結構化數據（比如圖像語音等）進項更好的識別、判斷和分類，讓演算法可以從數據和訓練中得到學習。這樣就像人腦一樣，只需要工程師透過類似的場景不斷對機器進行訓練，它就能自己學會做出判斷，這樣即使在全新的場景裡，車子也知道如何處理，更有利於適應和擴展。

Its differentiation from Intel， Google， Tesla is its lower cost or dependency on hardware. We can say drive.ai is software-driven driverless technology, and others are sensor-driven one.

基於深度學習的自動駕駛系統可以擺脫對於昂貴硬體的依賴。和特斯拉與Waymo的「天價」定制傳感器不同，Drive.ai使用的是商業化的低成本硬件，包括激光雷達、雷達和相機，深度學習系統會同步所有的傳感器數據，來基於這些信息作出最明智的決策，避免單個噪點導致的誤判。這樣即使其中一個失靈了，別的也可以正常工作。