{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"關於自動駕駛基礎架構"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"基礎架構在互聯網行業中，是一個相對比較成熟的領域。然而在自動駕駛領域，卻是一個新鮮的話題。基礎架構的工作包括硬件、onboard（車載系統）、雲端三大板塊。在我們認爲，"},{"type":"text","marks":[{"type":"strong"}],"text":"自動駕駛領域中 “基礎架構” 的核心價值，是爲自動駕駛提供恰到好處的、全方位的技術保障。在自動駕駛系統中，如果說感知是眼睛，規劃是大腦，那麼基礎架構就是神經系統，將自動駕駛軟件系統與車輛緊密的聯繫在一起。"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/d2\/d255d56eac046b6268d4eb8501d76fce.webp","alt":"Image","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"然而在近幾年的探索中，我們發現隨着自動駕駛技術的演進，種種基於安全、效率、體驗等方面的考慮也更加苛刻。基礎架構的工作不僅僅是爲軟件算法提供支持，還需要在硬件、效率、功能安全等多個方面進行權衡。而基礎架構工作本身，則是在三個主要技術矛盾之間進行權衡。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"技術的主要矛盾"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"▍1. 快速迭代與功能安全的矛盾"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"傳統的互聯網企業，其核心技術競爭力在於快速迭代。通過縮短迭代的週期，來加速功能的升級，進而在市場上獲取先發制人的競爭力。而傳統的汽車行業，則往往需要四至八年作爲一個迭代週期，以儘可能地確保功能安全。可見，“快速迭代” 和 “功能安全” 從時間的維度上看本身就是矛盾的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在自動駕駛領域，這個矛盾則更加突出。多數自動駕駛行業的從業者，可能都思考過這樣一個問題：在操作系統的選擇方面，是選擇更加熟悉、資源更加豐富的 Linux 系統，還是選擇實時性更有保障、更加安全可靠的 RTOS 系統？如果選擇前者，意味着可以更快地進行功能迭代，讓自動駕駛軟件算法快速適應複雜的交通場景，但很難保證極端情況下的功能安全；如果選擇後者，則意味着需要遵守更加嚴格的研發流程，需要做更多的系統性工作，而這本身就意味着時間。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"▍2. 更多、更好的傳感器與算力平臺的矛盾"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"坊間曾盛傳着一句話，在無限 ODD（設計行駛區域） 的場景下，感知對於傳感器的需求也是無限的。由近及遠的探測範圍、360° 的探測視角、多種類型傳感器的互補和冗餘，對於應對各種極端場景而言都是必不可少的。近幾年傳感器領域的發展也遵循着這個趨勢，Lidar 的點雲密度越來越大，Camera 的像素越來越高，自動駕駛廠商所使用的傳感器數量也越來越多。行業領頭羊 Waymo 在第五代車上所使用的傳感器數量更是多達 40 多個。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/68\/68bd83e0e5672f39c24bca8306eb6016.webp","alt":"Image","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"然而，傳感器方面提升的同時，也往往意味着對計算平臺的算力要求有着大幅提升，需要更強的算力來處理更多的數據。而更多的算力則通常意味着更大的功耗和散熱需求。也許在雲端計算中，我們可以假設算力、功耗和散熱等問題都是不存在的。然而在乘用車的狹小空間內，這個假設顯然不成立。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"▍3. 硬件性能與車規級安全的矛盾"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"基於 x86 平臺的小型服務器在搭上互聯網的順風車經過多年快速發展後，其計算能力是非常好的。然而問題在於，汽車本身並不是經過精心設計的 IDC（互聯網數據中心），對多個方面都有很強的約束，包括供電、散熱、體積、溫度範圍、電磁穩定、接口穩定、通信穩定等等多個方面。甚至在傳統汽車行業對這些條件都有明確的行業標準。這就意味着將傳統基於 x86 架構的工控機直接搬到車內是有很多問題的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"然而，即使經過多年的發展，滿足這些條件的計算平臺在算力方面卻往往不盡如人意。比如廣泛使用的 Nvidia Xavier 平臺，算力僅僅只有 30 TOPs。通過這樣的算力來實現全功能的 L4 級自動駕駛，幾乎是一件不可能的事情。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"同樣的，其他很多硬件比如交換機、傳感器皆是如此。在當前這個階段，工業級或者車規級產品的性能往往會更弱，這是硬件發展的客觀規律決定的。因此我們需要考慮的不僅僅是推動相關硬件行業的發展，還需要考慮軟硬件的協同演進，不能讓硬件成爲軟件算法的天花板。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"上述這三個矛盾，是當前階段我們在設計基礎架構時所面臨的三個核心問題。當然，"},{"type":"text","marks":[{"type":"strong"}],"text":"從自動駕駛的終極目標來看，功能安全、高算力平臺、穩定可靠的硬件都是不可或缺的。現階段我們探索更多的是，在資源有限的前提之下，如何用一種適當的節奏來逐步達成這個目標。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這樣就引申出另一個問題，自動駕駛的長遠目標是什麼？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"自動駕駛的長遠目標"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"自動駕駛的長期目標是用更安全穩定的無人駕駛，將人類駕駛員從重複，低效或危險（比如疲勞，醉酒，或極端天氣等）的駕駛行爲中解放出來，提升社會的運轉效率，降低交通運輸風險。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"運載的本質在於安全，沒有安全保障的無人駕駛，是對生命的漠視。那麼如何定義安全？這是自動駕駛領域內經久不衰的話題。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"百年汽車行業，不僅僅是百年的技術演進，更是持續百年的對汽車安全的探索。也正是如此，逐漸演變出 iso26262 等等非常全面的行業標準，對硬件、軟件、研發流程都有非常細緻的要求。這本質上是一種主動證明安全的方式，凡是按照車規級和功能安全研發的產品，至少可以保證是在一定的安全水準之上的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"而非汽車行業出身的玩家中，則逐漸形成另一種對於安全的定義。除了對每一個環節做到車規級的功能安全這種思路之外， 也可以通過更好的冗餘系統設計來提升整體安全性。比如當單車智能不足的時候，通過外部信息的幫助，比如 V2X、遠程接管等技術手段，來更快的實現安全前提下的無人駕駛。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這種方式，本質上可以認爲是對功能安全的定義的演進。被動層面上，可以通過 MPI （每兩次人工干預之間行駛的平均里程數）達到一個足夠大的數值，來證明整體系統的安全可靠。主動層面上，可以通過增加冗餘系統和降級系統，在各種故障場景下對於安全進行保障。除此之外，在軟件算法層面，也可以通過窮舉交通參與者的任何可能狀態和行爲，通過概率空間的方法，來證明軟件算法對各種異常場景的適應性。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/7c\/7c44646fb7772a05b28ee17a0bbd4e1f.webp","alt":"Image","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/df\/dfbd4cf060b6a7b02d9bc9b743eecf88.webp","alt":"Image","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/7b\/7b900a28da47268a1c62e8a3aa3ec1d1.webp","alt":"Image","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"雖然自動駕駛與汽車緊密相關，但畢竟自動駕駛是一個相對新興的行業，汽車行業對功能安全的定義和標準並不一定完全適用於自動駕駛領域。我們也許可以找到一個新的對於功能安全的定義和標準，來證明自動駕駛系統的安全可靠。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"也正是因爲這些原因，自動駕駛領域的玩家們，逐漸演變出兩種不同的流派："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"自下而上，安全爲重"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"自上而下，功能爲先"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這裏的 “上” 與 “下”，可以有三層含義："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在 SAE International 所制定的 “L2 L3 L4” 自動駕駛等級中，選擇相對複雜的 L4 爲起點，還是選擇相對簡單一些的 L2 爲起點。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在系統演進的過程中，優先選擇功能的實現，還是優先選擇系統的安全。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在覈心決策流程中，是以核心軟件算法爲主，由軟件定義車輛；還是以車輛硬件爲主，由硬件來約束軟件。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"可以看到，這與我們所遇到的基礎架構設計中的主要矛盾是相互對應的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"汽車行業出身的玩家，多采用第一種方式。嚴格遵循安全爲重的原則，在滿足車規級安全等級要求的前提下，選擇更加穩定可靠的硬件、系統和中間件，然後逐步進行功能實現。從這個角度講，汽車行業出身的玩家優先選擇 L2、L3 或者 ADAS 作爲自動駕駛的起點，並不完全是因爲更認可 L2、L3 的商業價值，更是因爲在這個前提之下，比較難實現安全可靠的 L4 級自動駕駛。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"非汽車行業出身的玩家，則多采用第二種方式。採用自上而下、功能爲先的研發方式，在有安全駕駛員的前提下，優先實現自動駕駛系統的核心功能，再逐步通過系統冗餘設計，以及 V2X、遠程接管等外部輔助手段，逐漸實現安全可靠的無人駕駛。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"滴滴自動駕駛從技術複雜性幾何級躍遷的 L4 級自動駕駛切入，鎖定 Robotaxi 的商業運營爲長遠目標，並在“終點“的指引下，不斷探索新的技術方案，同時兼顧軟件功能與安全，不斷積累能力，以一種適當的節奏，來逐步達成這個目標。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"自動駕駛的發展節奏"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在我認爲，如果以 MPI 爲參考系的話，整體上而言自動駕駛可以劃分爲四個階段："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"10 -> 100 -> 1000 -> driverless（無人駕駛）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"▍1. 第一階段：MPI < 10"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這個階段，可以定義爲自動駕駛的原型探索階段。如果按照每輛車每天可以測試 200km 里程的話，意味着每輛車每天可以產生 20 多個接管問題，少數的測試車輛即可滿足常見問題的穩定復現。因此，從這個角度看，在這個階段以無人駕駛爲目標是非常困難的。相對的，把安全駕駛員看作系統整體的一部分，更充分的利用人的存在來獲取更多的信息、數據並保證安全底線，反而顯得更有價值。在這個思路之下："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"在Infra 方面，"},{"type":"text","text":"其核心價值在於提供一個強有力的可快速迭代的系統架構和硬件平臺，賦能軟件算法的快速迭代，完成系統原型的搭建。在這個階段過於追求硬件、系統的穩定性和可靠性，不僅意義不大，還可能會對核心算法的探索造成阻礙。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"在Simulation 方面，"},{"type":"text","text":"由於日常路測會直接產生大量接管問題，對這些真實問題的高效分析和迭代是更重要的。因此對於仿真平臺而言，其核心價值並不在於構建更多的虛擬場景，而在於對軟件的迴歸測試和功能層面的白盒測試，而其他測試則交給封閉場地的結構化測試和開放道路的黑盒測試則相對更爲高效。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"在Autonomy 算法方面，"},{"type":"text","text":"對於如何實現自動駕駛的技術路徑應該還不足夠清晰和明確，需要大量的原型驗證和方法探索，並在這個過程中逐步形成相對明確的技術路線和方向，搭建一個完整軟件技術體系，形成數據驅動的研發流程。相對而言，一些頻次較低的問題和場景，以及一些比較極端安全問題，則可以優先級低一些，由安全員來進行保障。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"▍2. 第二階段：MPI <100 "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這個階段，可以定義爲自動駕駛的成長階段。這個階段，應該是方法論逐步成型並持續需求突破的一個過程。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"在 Infra 方面，"},{"type":"text","text":"需要對多樣化的系統架構和硬件平臺進行探索，給核心算法提供更好的傳感器和算力的支撐，幫助算法團隊進行持續深入的探索和更多方法的嘗試，賦能更多、硬件和安全的可能性。與此同時，由適當的力的約束，需要在系統層面逐步提供一些約束和優化，與算法團隊共同提升計算效能，從而在有限的資源下提供更多的可能性。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"在 Simulation 方面，"},{"type":"text","text":"需要提供更高效和低成本的測試手段，通過更好的仿真度和虛擬仿真，將部分需要依賴於實體車的場景測試和部分結構化測試轉移至仿真平臺，實現 Simulation In the Loop，提升研發測試的效率。從需求的角度講，如果將仿真分爲 “感知仿真” 和 “規劃仿真” 的話，則對於 “規劃仿真” 的仿真度的提升相對而言性價比更高。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"在 Autonomy 算法論、技術路徑方面，"},{"type":"text","text":"需要更多的傳感器和方法論方面的探索。一方面可以嘗試更好的傳感器和硬件所帶來的直接收益，另一方面，則需要在方法論方面有一些突破，能夠更全面的來應對更復雜的道路場景。於此同時，由於計算平臺算力的約束，也需要在計算效能方面有更多的嘗試，在算力有限的情況下嘗試更多的可能性。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"▍3. 第三階段：MPI < 1000"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這個階段，可以定義爲自動駕駛的蛻變階段。這個階段會有兩個特徵："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"自動駕駛的 issue 逐漸進入長尾階段，issue 的發現和處理效率將會成爲核心瓶頸"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"系統穩定性和可靠性將會逐步成爲 MPI 的主要組成部分"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"所以，對於自動駕駛系統而言，需要向功能安全做出更多的嘗試和探索，追求系統的穩定性和可靠性。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"在 Infra 方面，"},{"type":"text","text":"需要全面向安全可靠的系統架構進行探索。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"硬件方面，需要向車規級傳感器、高能耗比計算平臺和連接方式靠攏，探索高能耗比專用芯片的可能性；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"中間件方面，對上要對算法模塊提供更加深入的約束和頂層設計，來實現更好的系統穩定性和可靠性；對下要兼容不同架構的硬件平臺和操作系統，提供對車規級硬件的適配能力；對內要實現可靠的故障探測機制和監控系統，作爲承上啓下的核心環節；對外要提供遠程監控、遠程協助甚至遠程接管的能力，提供全方位的安全保障。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"系統方面，需要進行充分的系統性冗餘設計，通過倒金字塔式的收斂方法探索不同層級的 failover 能力，實現整體系統的安全可靠。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"在 Simulation 方面，"},{"type":"text","text":"需要提升發現問題的效率，減少對真實路測的依賴。可以嘗試一些諸如對交通參與者的行爲進行窮舉的方法，從所有可能性中找出 In Scope 的範圍和邊界，來驗證自動駕駛算法對於所有可能性的應對能力。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"在 Autonomy 算法方面，"},{"type":"text","text":"需要在算法本身相對成熟的前提下，通過軟硬件結合的方式進行系統性的優化，提升算法的運行效率、穩定性和可靠性。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"綜上所述，"},{"type":"text","marks":[{"type":"strong"}],"text":"以 MPI 爲參考系來劃分多個階段的核心價值，在於通過這種方式來制定一個更加清晰的 Roadmap，明確每個階段的核心目標。從而在總體資源有限的前提之下，步步爲營地朝着自動駕駛的終極目標演進。同樣，這對基礎架構的演進節奏和目標設定是有很好的參考意義的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"基礎架構該如何演進"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"作爲自動駕駛整個系統的基礎設施，基礎架構相關的工作需要比軟件算法模塊更早一步。這也是所謂的 Infra 先行。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"從當前行業所處的階段並結合國內道路的實際情況來看，在自動駕駛這條長跑賽道上，多數玩家仍然處於相對早期的階段。我們需要清晰的認識到，實現普遍意義上的無人駕駛還需要較長的時間。因此，我們可以假設無人駕駛的落地更可能是一種由點及面的過程，那麼特定區域、有限 ODD（設計運行區域）之內的無人駕駛商業化落地，將會是這個行業未來重要的里程碑，正如 Waymo 在鳳凰城做的那樣。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在這個目標之下，以 MPI 爲參考系，基礎架構的演進需要將重點放在兩個方面："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"硬件架構升級：從通用的硬件體系逐步向高性能的、穩定可靠的專用硬件體系演進。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"軟件架構升級：從賦能算法逐步向自上而下的整體軟件架構設計演進，實現更加安全的、全面的基礎軟件體系。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"▍1. 硬件架構的演進思路"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"自動駕駛的硬件部分，通常包括傳感器和計算平臺兩個部分。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"1.1 傳感器部分"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"從行業的發展趨勢來看，除了極少數玩家，如 Waymo，投入大量資源用於傳感器的自研以外，已經逐步形成一個相對完整的供應商體系，"},{"type":"text","marks":[{"type":"strong"}],"text":"在保持性能提升的同時，逐步朝着車規級的目標演進。"},{"type":"text","text":"這裏對我們影響較大的主要有兩種類型的傳感器："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"Lidar:"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"追求更高的點雲密度和更遠的探測距離，並逐步車規化。通信方面，由傳統的 Ethernet，逐步向車載以太網演進，並在硬件上逐漸實現車規級。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"Camera："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"追求更高的分辨率和更大的動態範圍，適配不同光線的場景。通信方面，由工業級的 USB、Ethernet 接口逐步向高帶寬的串行數據接口（比如 GMSL based LVDS 協議）演進，以獲取更大的通信帶寬、更低的傳輸延時和更好的接插件可靠性。不過由此引入的問題是，車規級 Camera 通常由於體積的約束，多數需要 ISP 外置，這意味着我們可能要涉及更多與 ISP 相關的工作。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"由此可見，在傳感器方面，能施展的空間並不大。我們可能更需要關注的是，如何能夠結合感知模塊的需求與平臺算力的約束，更加充分的利用這些傳感器的特性，爲自動駕駛的眼睛提供更多的信息。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"1.2 計算平臺部分"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這是相關問題的核心所在，也是基礎架構團隊需要解決的核心矛盾。多數互聯網出身的自動駕駛團隊，都是從一臺工控機開始的，使用基於 Intel CPU + Nvidia GPU 的解決方案作爲計算平臺。然而作爲運行在車上的計算平臺，在硬件本身方面，我們至少需要考慮："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"供電"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"車輛平臺尤其是電動汽車多提供 12V 的直流供電方式。相對於傳統的 220V 交流供電而言，如果可以適配這種供電方式，不僅可以增加效能，也能避免一些安全隱患。除此之外，車輛供電所能支持的最大功率也是一個很重要的問題。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"散熱"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"汽車上可以使用的散熱方式主要有風冷和液冷兩種。風冷雖然成本比較低，結構簡單，不過缺點在於散熱效率較低，噪音也比較大。與此同時，多數純電動汽車或者混動汽車，都會有原生的電池液冷迴路。因此，相對而言，液冷方案更加適用於車載環境。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"接口"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Ethernet 是階段非常常用的一種大量數據的傳輸方式。然而傳統 Ethernet 多采用 RJ45 的接口形式，在長期運行的過程中會對網絡的穩定性產生一定影響。除此之外，也需要足夠多的接口能夠接入所需的傳感器。除此之外，CAN\/串口等等也是比較常用的接口類型。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"從當前的行業現狀來看，如果希望達到這些目標，可以有兩種途徑："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"一、採購相對成熟的解決方案。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"比如 Nvidia 的 Xavier 以及基於 Xavier 的 Pegasus 平臺。Nvidia 依靠在 Deep Learning 領域內的深厚積累，已經在自動駕駛領域深耕多年，並有望成爲一個新的 Tier 1 勢力。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Xavier 或者 Pegasus 的優勢主要有："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"是一個完整的系統，並且是按照車規級的標準設計的；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"已經適配了很多傳感器，尤其是提供了 ISP 適配了很多 Camera；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"有着相對清晰的 Roadmap，方便後續的迭代升級。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"然而核心問題在於 SoC 中的 CPU 部分性能太弱。這顆基於 Arm 架構的 CPU，主頻爲 1.8GHz，8 個核心，難以滿足核心算法的性能需求。也許這也是 Nvidia 選擇收購 Arm 的原因之一，在未來可以逐步補全 CPU 方面的短板。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"二、採用自己主導設計，定製化的解決方案。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"以 x86 平臺爲原型，直接集成 Intel Xeon 的 CPU 與 Nvidia 最新架構的 GPU，以期達到最強算力。不過這個方案的缺點在於，除了需要對接口、散熱、電源進行定製以外，還需要一些與傳感器集成相關的工作，以及未來需要克服更多的困難來往車規級和功能安全演進。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這兩種方案，本質上也是車規級方案與性能向方案的抉擇。現實就是這樣，工業級（航空級、車規級）產品相對而言雖然更加穩定可靠，但性能往往要比普通的消費級產品差一些。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我們再回到 MPI 這個參照系上。在前兩個階段中，如果選用諸如 Nvidia Pegasus 這樣的方案，雖然在可靠性和傳感器接入方面有一定的優勢，但性能方面的問題將會在一定程度上制約上層軟件算法的迭代和更多方法的嘗試。由此來看，在這兩個方案之間選擇更加安全可靠但是性能相對較低的計算平臺，並不是一個很好的選擇。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"不過我們需要看到，在當前狀態下，Nvidia 在自動駕駛及相關領域的投入是遠比 Intel 等其他廠商要大的，也取得了一定的成績。如果可以做到補齊 CPU 方面的短板，或者在軟件方面減少對於 CPU 的依賴，那麼未來大概率還是會朝着 Nvidia Pegasus 類似的方向上演進。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"▍2. 軟件架構的演進思路"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"自動駕駛的軟件架構，通常包括操作系統、中間件、監控系統等等多個部分。中間件則是軟件架構中最核心的部分，介於車輛硬件、計算平臺與上層軟件算法之間，起到承上啓下的作用。在中間件部分，我們可能耳熟能詳的有："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"ROS"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Apollo Cyber RT"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Iceoryx"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這三者之間的區別就不在這裏詳細闡述了。不過從某種程度上來看，這三者也分別代表了三種不同的設計理念："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"ROS：設計成爲一個相對通用的分佈式機器人控制平臺。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Apollo Cyber RT：設計爲一個高性能的、全功能的自動駕駛專用平臺。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Iceoryx：設計爲一個高性能的、跨平框架功能安全的自動駕駛通信框 。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這三種設計理念，也是與上文所述的三個階段隱隱有所呼應的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"多數自動駕駛領域的玩家，都是從 ROS\/Ubuntu 開始，主要原因在於其成熟、完整，技術人員可以更多的 focus 在自動駕駛的核心算法上。然後隨着項目的進展，逐漸沿着 Apollo Cyber RT 類似的理念進行演進，針對自動駕駛的具體場景進行技術優化，逐步形成更加完整、高效的軟件架構。最終朝着功能安全的方向發展，逐步實現量產化。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這裏面的核心問題，在於 “靈活” 與 “可控” 的權衡。正如要蓋一棟大樓，可以選擇建造只有主體框架結構的商業建築，由用戶自行設計內部空間；也可以選擇建造具備戶型設計的住宅建築，用戶只能在相應功能區域之內做設計。目的不同，設計理念自然不同。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"因此，中間件的設計，不能一味的追求功能，而應該結合當前軟件與硬件的發展階段，考慮所需解決的核心問題與階段性項目目標，來進行整體的設計和規劃。一個好的中間件設計，需要考慮整體軟件架構中的“靈活”與“可控”，需要考慮不同硬件架構和操作系統的適配，需要權衡功能安全與軟件效率，需要能夠賦能整個項目的研發與迭代。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"結語"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"自動駕駛是一個非常複雜的系統，包含感知、預測、規劃、控制、地圖、仿真、車輛、基礎架構等等多個技術方向。這些方向之間並不僅僅只是“木桶理論”中的長短的關係，而更是一種相互依存、相互制約、協同發展的關係。正如兩人三足遊戲中，個體的激進只會造成摔跤，整體的步伐一致、節奏相符纔可以走的更快。因此，如何理解自動駕駛的發展節奏，如何在各個方向上實現步伐一致並與整體節奏相符，是每個方向都需要考慮的問題，更是基礎架構方向演進的基礎。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"自動駕駛是一個顛覆性的創新技術，不可否認仍然有很多未解的技術難題需要去攻克。但是值得慶幸地是，整個行業都在堅定地全力推進，相信技術驅動改變世界的那一天並不遙遠。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"horizontalrule"},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"頭圖：Unsplash"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"作者：Ferry Wang"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"原文：https:\/\/mp.weixin.qq.com\/s\/6CTKIdqupCPouyRHdBM-EQ"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"原文：滴滴自動駕駛 | 基礎架構如何以終爲始，穩定先行？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"來源：滴滴技術 - 微信公衆號 [ID：didi_tech]"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"轉載：著作權歸作者所有。商業轉載請聯繫作者獲得授權，非商業轉載請註明出處。"}]}]}
基礎架構如何以終爲始，穩定先行？
{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"關於自動駕駛基礎架構"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"基礎架構在互聯網行業中，是一個相對比較成熟的領域。然而在自動駕駛領域，卻是一個新鮮的話題。基礎架構的工作包括硬件、onboard（車載系統）、雲端三大板塊。在我們認爲，"},{"type":"text","marks":[{"type":"strong"}],"text":"自動駕駛領域中 “基礎架構” 的核心價值，是爲自動駕駛提供恰到好處的、全方位的技術保障。在自動駕駛系統中，如果說感知是眼睛，規劃是大腦，那麼基礎架構就是神經系統，將自動駕駛軟件系統與車輛緊密的聯繫在一起。"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/d2\/d255d56eac046b6268d4eb8501d76fce.webp","alt":"Image","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"然而在近幾年的探索中，我們發現隨着自動駕駛技術的演進，種種基於安全、效率、體驗等方面的考慮也更加苛刻。基礎架構的工作不僅僅是爲軟件算法提供支持，還需要在硬件、效率、功能安全等多個方面進行權衡。而基礎架構工作本身，則是在三個主要技術矛盾之間進行權衡。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"技術的主要矛盾"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"▍1. 快速迭代與功能安全的矛盾"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"傳統的互聯網企業，其核心技術競爭力在於快速迭代。通過縮短迭代的週期，來加速功能的升級，進而在市場上獲取先發制人的競爭力。而傳統的汽車行業，則往往需要四至八年作爲一個迭代週期，以儘可能地確保功能安全。可見，“快速迭代” 和 “功能安全” 從時間的維度上看本身就是矛盾的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在自動駕駛領域，這個矛盾則更加突出。多數自動駕駛行業的從業者，可能都思考過這樣一個問題：在操作系統的選擇方面，是選擇更加熟悉、資源更加豐富的 Linux 系統，還是選擇實時性更有保障、更加安全可靠的 RTOS 系統？如果選擇前者，意味着可以更快地進行功能迭代，讓自動駕駛軟件算法快速適應複雜的交通場景，但很難保證極端情況下的功能安全；如果選擇後者，則意味着需要遵守更加嚴格的研發流程，需要做更多的系統性工作，而這本身就意味着時間。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"▍2. 更多、更好的傳感器與算力平臺的矛盾"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"坊間曾盛傳着一句話，在無限 ODD（設計行駛區域） 的場景下，感知對於傳感器的需求也是無限的。由近及遠的探測範圍、360° 的探測視角、多種類型傳感器的互補和冗餘，對於應對各種極端場景而言都是必不可少的。近幾年傳感器領域的發展也遵循着這個趨勢，Lidar 的點雲密度越來越大，Camera 的像素越來越高，自動駕駛廠商所使用的傳感器數量也越來越多。行業領頭羊 Waymo 在第五代車上所使用的傳感器數量更是多達 40 多個。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/68\/68bd83e0e5672f39c24bca8306eb6016.webp","alt":"Image","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"然而，傳感器方面提升的同時，也往往意味着對計算平臺的算力要求有着大幅提升，需要更強的算力來處理更多的數據。而更多的算力則通常意味着更大的功耗和散熱需求。也許在雲端計算中，我們可以假設算力、功耗和散熱等問題都是不存在的。然而在乘用車的狹小空間內，這個假設顯然不成立。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"▍3. 硬件性能與車規級安全的矛盾"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"基於 x86 平臺的小型服務器在搭上互聯網的順風車經過多年快速發展後，其計算能力是非常好的。然而問題在於，汽車本身並不是經過精心設計的 IDC（互聯網數據中心），對多個方面都有很強的約束，包括供電、散熱、體積、溫度範圍、電磁穩定、接口穩定、通信穩定等等多個方面。甚至在傳統汽車行業對這些條件都有明確的行業標準。這就意味着將傳統基於 x86 架構的工控機直接搬到車內是有很多問題的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"然而，即使經過多年的發展，滿足這些條件的計算平臺在算力方面卻往往不盡如人意。比如廣泛使用的 Nvidia Xavier 平臺，算力僅僅只有 30 TOPs。通過這樣的算力來實現全功能的 L4 級自動駕駛，幾乎是一件不可能的事情。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"同樣的，其他很多硬件比如交換機、傳感器皆是如此。在當前這個階段，工業級或者車規級產品的性能往往會更弱，這是硬件發展的客觀規律決定的。因此我們需要考慮的不僅僅是推動相關硬件行業的發展，還需要考慮軟硬件的協同演進，不能讓硬件成爲軟件算法的天花板。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"上述這三個矛盾，是當前階段我們在設計基礎架構時所面臨的三個核心問題。當然，"},{"type":"text","marks":[{"type":"strong"}],"text":"從自動駕駛的終極目標來看，功能安全、高算力平臺、穩定可靠的硬件都是不可或缺的。現階段我們探索更多的是，在資源有限的前提之下，如何用一種適當的節奏來逐步達成這個目標。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這樣就引申出另一個問題，自動駕駛的長遠目標是什麼？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"自動駕駛的長遠目標"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"自動駕駛的長期目標是用更安全穩定的無人駕駛，將人類駕駛員從重複，低效或危險（比如疲勞，醉酒，或極端天氣等）的駕駛行爲中解放出來，提升社會的運轉效率，降低交通運輸風險。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"運載的本質在於安全，沒有安全保障的無人駕駛，是對生命的漠視。那麼如何定義安全？這是自動駕駛領域內經久不衰的話題。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"百年汽車行業，不僅僅是百年的技術演進，更是持續百年的對汽車安全的探索。也正是如此，逐漸演變出 iso26262 等等非常全面的行業標準，對硬件、軟件、研發流程都有非常細緻的要求。這本質上是一種主動證明安全的方式，凡是按照車規級和功能安全研發的產品，至少可以保證是在一定的安全水準之上的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"而非汽車行業出身的玩家中，則逐漸形成另一種對於安全的定義。除了對每一個環節做到車規級的功能安全這種思路之外， 也可以通過更好的冗餘系統設計來提升整體安全性。比如當單車智能不足的時候，通過外部信息的幫助，比如 V2X、遠程接管等技術手段，來更快的實現安全前提下的無人駕駛。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這種方式，本質上可以認爲是對功能安全的定義的演進。被動層面上，可以通過 MPI （每兩次人工干預之間行駛的平均里程數）達到一個足夠大的數值，來證明整體系統的安全可靠。主動層面上，可以通過增加冗餘系統和降級系統，在各種故障場景下對於安全進行保障。除此之外，在軟件算法層面，也可以通過窮舉交通參與者的任何可能狀態和行爲，通過概率空間的方法，來證明軟件算法對各種異常場景的適應性。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/7c\/7c44646fb7772a05b28ee17a0bbd4e1f.webp","alt":"Image","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/df\/dfbd4cf060b6a7b02d9bc9b743eecf88.webp","alt":"Image","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/7b\/7b900a28da47268a1c62e8a3aa3ec1d1.webp","alt":"Image","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"雖然自動駕駛與汽車緊密相關，但畢竟自動駕駛是一個相對新興的行業，汽車行業對功能安全的定義和標準並不一定完全適用於自動駕駛領域。我們也許可以找到一個新的對於功能安全的定義和標準，來證明自動駕駛系統的安全可靠。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"也正是因爲這些原因，自動駕駛領域的玩家們，逐漸演變出兩種不同的流派："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"自下而上，安全爲重"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"自上而下，功能爲先"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這裏的 “上” 與 “下”，可以有三層含義："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在 SAE International 所制定的 “L2 L3 L4” 自動駕駛等級中，選擇相對複雜的 L4 爲起點，還是選擇相對簡單一些的 L2 爲起點。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在系統演進的過程中，優先選擇功能的實現，還是優先選擇系統的安全。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在覈心決策流程中，是以核心軟件算法爲主，由軟件定義車輛；還是以車輛硬件爲主，由硬件來約束軟件。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"可以看到，這與我們所遇到的基礎架構設計中的主要矛盾是相互對應的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"汽車行業出身的玩家，多采用第一種方式。嚴格遵循安全爲重的原則，在滿足車規級安全等級要求的前提下，選擇更加穩定可靠的硬件、系統和中間件，然後逐步進行功能實現。從這個角度講，汽車行業出身的玩家優先選擇 L2、L3 或者 ADAS 作爲自動駕駛的起點，並不完全是因爲更認可 L2、L3 的商業價值，更是因爲在這個前提之下，比較難實現安全可靠的 L4 級自動駕駛。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"非汽車行業出身的玩家，則多采用第二種方式。採用自上而下、功能爲先的研發方式，在有安全駕駛員的前提下，優先實現自動駕駛系統的核心功能，再逐步通過系統冗餘設計，以及 V2X、遠程接管等外部輔助手段，逐漸實現安全可靠的無人駕駛。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"滴滴自動駕駛從技術複雜性幾何級躍遷的 L4 級自動駕駛切入，鎖定 Robotaxi 的商業運營爲長遠目標，並在“終點“的指引下，不斷探索新的技術方案，同時兼顧軟件功能與安全，不斷積累能力，以一種適當的節奏，來逐步達成這個目標。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"自動駕駛的發展節奏"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在我認爲，如果以 MPI 爲參考系的話，整體上而言自動駕駛可以劃分爲四個階段："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"10 -> 100 -> 1000 -> driverless（無人駕駛）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"▍1. 第一階段：MPI < 10"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這個階段，可以定義爲自動駕駛的原型探索階段。如果按照每輛車每天可以測試 200km 里程的話，意味着每輛車每天可以產生 20 多個接管問題，少數的測試車輛即可滿足常見問題的穩定復現。因此，從這個角度看，在這個階段以無人駕駛爲目標是非常困難的。相對的，把安全駕駛員看作系統整體的一部分，更充分的利用人的存在來獲取更多的信息、數據並保證安全底線，反而顯得更有價值。在這個思路之下："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"在Infra 方面，"},{"type":"text","text":"其核心價值在於提供一個強有力的可快速迭代的系統架構和硬件平臺，賦能軟件算法的快速迭代，完成系統原型的搭建。在這個階段過於追求硬件、系統的穩定性和可靠性，不僅意義不大，還可能會對核心算法的探索造成阻礙。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"在Simulation 方面，"},{"type":"text","text":"由於日常路測會直接產生大量接管問題，對這些真實問題的高效分析和迭代是更重要的。因此對於仿真平臺而言，其核心價值並不在於構建更多的虛擬場景，而在於對軟件的迴歸測試和功能層面的白盒測試，而其他測試則交給封閉場地的結構化測試和開放道路的黑盒測試則相對更爲高效。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"在Autonomy 算法方面，"},{"type":"text","text":"對於如何實現自動駕駛的技術路徑應該還不足夠清晰和明確，需要大量的原型驗證和方法探索，並在這個過程中逐步形成相對明確的技術路線和方向，搭建一個完整軟件技術體系，形成數據驅動的研發流程。相對而言，一些頻次較低的問題和場景，以及一些比較極端安全問題，則可以優先級低一些，由安全員來進行保障。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"▍2. 第二階段：MPI <100 "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這個階段，可以定義爲自動駕駛的成長階段。這個階段，應該是方法論逐步成型並持續需求突破的一個過程。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"在 Infra 方面，"},{"type":"text","text":"需要對多樣化的系統架構和硬件平臺進行探索，給核心算法提供更好的傳感器和算力的支撐，幫助算法團隊進行持續深入的探索和更多方法的嘗試，賦能更多、硬件和安全的可能性。與此同時，由適當的力的約束，需要在系統層面逐步提供一些約束和優化，與算法團隊共同提升計算效能，從而在有限的資源下提供更多的可能性。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"在 Simulation 方面，"},{"type":"text","text":"需要提供更高效和低成本的測試手段，通過更好的仿真度和虛擬仿真，將部分需要依賴於實體車的場景測試和部分結構化測試轉移至仿真平臺，實現 Simulation In the Loop，提升研發測試的效率。從需求的角度講，如果將仿真分爲 “感知仿真” 和 “規劃仿真” 的話，則對於 “規劃仿真” 的仿真度的提升相對而言性價比更高。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"在 Autonomy 算法論、技術路徑方面，"},{"type":"text","text":"需要更多的傳感器和方法論方面的探索。一方面可以嘗試更好的傳感器和硬件所帶來的直接收益，另一方面，則需要在方法論方面有一些突破，能夠更全面的來應對更復雜的道路場景。於此同時，由於計算平臺算力的約束，也需要在計算效能方面有更多的嘗試，在算力有限的情況下嘗試更多的可能性。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"▍3. 第三階段：MPI < 1000"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這個階段，可以定義爲自動駕駛的蛻變階段。這個階段會有兩個特徵："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"自動駕駛的 issue 逐漸進入長尾階段，issue 的發現和處理效率將會成爲核心瓶頸"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"系統穩定性和可靠性將會逐步成爲 MPI 的主要組成部分"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"所以，對於自動駕駛系統而言，需要向功能安全做出更多的嘗試和探索，追求系統的穩定性和可靠性。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"在 Infra 方面，"},{"type":"text","text":"需要全面向安全可靠的系統架構進行探索。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"硬件方面，需要向車規級傳感器、高能耗比計算平臺和連接方式靠攏，探索高能耗比專用芯片的可能性；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"中間件方面，對上要對算法模塊提供更加深入的約束和頂層設計，來實現更好的系統穩定性和可靠性；對下要兼容不同架構的硬件平臺和操作系統，提供對車規級硬件的適配能力；對內要實現可靠的故障探測機制和監控系統，作爲承上啓下的核心環節；對外要提供遠程監控、遠程協助甚至遠程接管的能力，提供全方位的安全保障。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"系統方面，需要進行充分的系統性冗餘設計，通過倒金字塔式的收斂方法探索不同層級的 failover 能力，實現整體系統的安全可靠。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"在 Simulation 方面，"},{"type":"text","text":"需要提升發現問題的效率，減少對真實路測的依賴。可以嘗試一些諸如對交通參與者的行爲進行窮舉的方法，從所有可能性中找出 In Scope 的範圍和邊界，來驗證自動駕駛算法對於所有可能性的應對能力。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"在 Autonomy 算法方面，"},{"type":"text","text":"需要在算法本身相對成熟的前提下，通過軟硬件結合的方式進行系統性的優化，提升算法的運行效率、穩定性和可靠性。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"綜上所述，"},{"type":"text","marks":[{"type":"strong"}],"text":"以 MPI 爲參考系來劃分多個階段的核心價值，在於通過這種方式來制定一個更加清晰的 Roadmap，明確每個階段的核心目標。從而在總體資源有限的前提之下，步步爲營地朝着自動駕駛的終極目標演進。同樣，這對基礎架構的演進節奏和目標設定是有很好的參考意義的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"基礎架構該如何演進"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"作爲自動駕駛整個系統的基礎設施，基礎架構相關的工作需要比軟件算法模塊更早一步。這也是所謂的 Infra 先行。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"從當前行業所處的階段並結合國內道路的實際情況來看，在自動駕駛這條長跑賽道上，多數玩家仍然處於相對早期的階段。我們需要清晰的認識到，實現普遍意義上的無人駕駛還需要較長的時間。因此，我們可以假設無人駕駛的落地更可能是一種由點及面的過程，那麼特定區域、有限 ODD（設計運行區域）之內的無人駕駛商業化落地，將會是這個行業未來重要的里程碑，正如 Waymo 在鳳凰城做的那樣。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在這個目標之下，以 MPI 爲參考系，基礎架構的演進需要將重點放在兩個方面："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"硬件架構升級：從通用的硬件體系逐步向高性能的、穩定可靠的專用硬件體系演進。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"軟件架構升級：從賦能算法逐步向自上而下的整體軟件架構設計演進，實現更加安全的、全面的基礎軟件體系。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"▍1. 硬件架構的演進思路"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"自動駕駛的硬件部分，通常包括傳感器和計算平臺兩個部分。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"1.1 傳感器部分"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"從行業的發展趨勢來看，除了極少數玩家，如 Waymo，投入大量資源用於傳感器的自研以外，已經逐步形成一個相對完整的供應商體系，"},{"type":"text","marks":[{"type":"strong"}],"text":"在保持性能提升的同時，逐步朝着車規級的目標演進。"},{"type":"text","text":"這裏對我們影響較大的主要有兩種類型的傳感器："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"Lidar:"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"追求更高的點雲密度和更遠的探測距離，並逐步車規化。通信方面，由傳統的 Ethernet，逐步向車載以太網演進，並在硬件上逐漸實現車規級。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"Camera："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"追求更高的分辨率和更大的動態範圍，適配不同光線的場景。通信方面，由工業級的 USB、Ethernet 接口逐步向高帶寬的串行數據接口（比如 GMSL based LVDS 協議）演進，以獲取更大的通信帶寬、更低的傳輸延時和更好的接插件可靠性。不過由此引入的問題是，車規級 Camera 通常由於體積的約束，多數需要 ISP 外置，這意味着我們可能要涉及更多與 ISP 相關的工作。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"由此可見，在傳感器方面，能施展的空間並不大。我們可能更需要關注的是，如何能夠結合感知模塊的需求與平臺算力的約束，更加充分的利用這些傳感器的特性，爲自動駕駛的眼睛提供更多的信息。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"1.2 計算平臺部分"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這是相關問題的核心所在，也是基礎架構團隊需要解決的核心矛盾。多數互聯網出身的自動駕駛團隊，都是從一臺工控機開始的，使用基於 Intel CPU + Nvidia GPU 的解決方案作爲計算平臺。然而作爲運行在車上的計算平臺，在硬件本身方面，我們至少需要考慮："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"供電"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"車輛平臺尤其是電動汽車多提供 12V 的直流供電方式。相對於傳統的 220V 交流供電而言，如果可以適配這種供電方式，不僅可以增加效能，也能避免一些安全隱患。除此之外，車輛供電所能支持的最大功率也是一個很重要的問題。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"散熱"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"汽車上可以使用的散熱方式主要有風冷和液冷兩種。風冷雖然成本比較低，結構簡單，不過缺點在於散熱效率較低，噪音也比較大。與此同時，多數純電動汽車或者混動汽車，都會有原生的電池液冷迴路。因此，相對而言，液冷方案更加適用於車載環境。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"接口"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Ethernet 是階段非常常用的一種大量數據的傳輸方式。然而傳統 Ethernet 多采用 RJ45 的接口形式，在長期運行的過程中會對網絡的穩定性產生一定影響。除此之外，也需要足夠多的接口能夠接入所需的傳感器。除此之外，CAN\/串口等等也是比較常用的接口類型。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"從當前的行業現狀來看，如果希望達到這些目標，可以有兩種途徑："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"一、採購相對成熟的解決方案。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"比如 Nvidia 的 Xavier 以及基於 Xavier 的 Pegasus 平臺。Nvidia 依靠在 Deep Learning 領域內的深厚積累，已經在自動駕駛領域深耕多年，並有望成爲一個新的 Tier 1 勢力。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Xavier 或者 Pegasus 的優勢主要有："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"是一個完整的系統，並且是按照車規級的標準設計的；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"已經適配了很多傳感器，尤其是提供了 ISP 適配了很多 Camera；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"有着相對清晰的 Roadmap，方便後續的迭代升級。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"然而核心問題在於 SoC 中的 CPU 部分性能太弱。這顆基於 Arm 架構的 CPU，主頻爲 1.8GHz，8 個核心，難以滿足核心算法的性能需求。也許這也是 Nvidia 選擇收購 Arm 的原因之一，在未來可以逐步補全 CPU 方面的短板。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"二、採用自己主導設計，定製化的解決方案。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"以 x86 平臺爲原型，直接集成 Intel Xeon 的 CPU 與 Nvidia 最新架構的 GPU，以期達到最強算力。不過這個方案的缺點在於，除了需要對接口、散熱、電源進行定製以外，還需要一些與傳感器集成相關的工作，以及未來需要克服更多的困難來往車規級和功能安全演進。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這兩種方案，本質上也是車規級方案與性能向方案的抉擇。現實就是這樣，工業級（航空級、車規級）產品相對而言雖然更加穩定可靠，但性能往往要比普通的消費級產品差一些。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我們再回到 MPI 這個參照系上。在前兩個階段中，如果選用諸如 Nvidia Pegasus 這樣的方案，雖然在可靠性和傳感器接入方面有一定的優勢，但性能方面的問題將會在一定程度上制約上層軟件算法的迭代和更多方法的嘗試。由此來看，在這兩個方案之間選擇更加安全可靠但是性能相對較低的計算平臺，並不是一個很好的選擇。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"不過我們需要看到，在當前狀態下，Nvidia 在自動駕駛及相關領域的投入是遠比 Intel 等其他廠商要大的，也取得了一定的成績。如果可以做到補齊 CPU 方面的短板，或者在軟件方面減少對於 CPU 的依賴，那麼未來大概率還是會朝着 Nvidia Pegasus 類似的方向上演進。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"▍2. 軟件架構的演進思路"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"自動駕駛的軟件架構，通常包括操作系統、中間件、監控系統等等多個部分。中間件則是軟件架構中最核心的部分，介於車輛硬件、計算平臺與上層軟件算法之間，起到承上啓下的作用。在中間件部分，我們可能耳熟能詳的有："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"ROS"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Apollo Cyber RT"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Iceoryx"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這三者之間的區別就不在這裏詳細闡述了。不過從某種程度上來看，這三者也分別代表了三種不同的設計理念："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"ROS：設計成爲一個相對通用的分佈式機器人控制平臺。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Apollo Cyber RT：設計爲一個高性能的、全功能的自動駕駛專用平臺。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Iceoryx：設計爲一個高性能的、跨平框架功能安全的自動駕駛通信框 。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這三種設計理念，也是與上文所述的三個階段隱隱有所呼應的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"多數自動駕駛領域的玩家，都是從 ROS\/Ubuntu 開始，主要原因在於其成熟、完整，技術人員可以更多的 focus 在自動駕駛的核心算法上。然後隨着項目的進展，逐漸沿着 Apollo Cyber RT 類似的理念進行演進，針對自動駕駛的具體場景進行技術優化，逐步形成更加完整、高效的軟件架構。最終朝着功能安全的方向發展，逐步實現量產化。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這裏面的核心問題，在於 “靈活” 與 “可控” 的權衡。正如要蓋一棟大樓，可以選擇建造只有主體框架結構的商業建築，由用戶自行設計內部空間；也可以選擇建造具備戶型設計的住宅建築，用戶只能在相應功能區域之內做設計。目的不同，設計理念自然不同。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"因此，中間件的設計，不能一味的追求功能，而應該結合當前軟件與硬件的發展階段，考慮所需解決的核心問題與階段性項目目標，來進行整體的設計和規劃。一個好的中間件設計，需要考慮整體軟件架構中的“靈活”與“可控”，需要考慮不同硬件架構和操作系統的適配，需要權衡功能安全與軟件效率，需要能夠賦能整個項目的研發與迭代。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"結語"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"自動駕駛是一個非常複雜的系統，包含感知、預測、規劃、控制、地圖、仿真、車輛、基礎架構等等多個技術方向。這些方向之間並不僅僅只是“木桶理論”中的長短的關係，而更是一種相互依存、相互制約、協同發展的關係。正如兩人三足遊戲中，個體的激進只會造成摔跤，整體的步伐一致、節奏相符纔可以走的更快。因此，如何理解自動駕駛的發展節奏，如何在各個方向上實現步伐一致並與整體節奏相符，是每個方向都需要考慮的問題，更是基礎架構方向演進的基礎。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"自動駕駛是一個顛覆性的創新技術，不可否認仍然有很多未解的技術難題需要去攻克。但是值得慶幸地是，整個行業都在堅定地全力推進，相信技術驅動改變世界的那一天並不遙遠。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"horizontalrule"},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"頭圖：Unsplash"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"作者：Ferry Wang"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"原文：https:\/\/mp.weixin.qq.com\/s\/6CTKIdqupCPouyRHdBM-EQ"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"原文：滴滴自動駕駛 | 基礎架構如何以終爲始，穩定先行？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"來源：滴滴技術 - 微信公衆號 [ID：didi_tech]"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"轉載：著作權歸作者所有。商業轉載請聯繫作者獲得授權，非商業轉載請註明出處。"}]}]}
發表評論
所有評論
還沒有人評論，想成為第一個評論的人麼? 請在上方評論欄輸入並且點擊發布.
相關文章