作者chandler0227 (錢德勒)
看板car
標題Re: [情報] Cruise L4自駕配備人類駕駛
時間Tue May 14 18:20:24 2024
[自駕]
1. 兩套以上獨立的自駕系統(備援、冗餘)
2. 承1，兩套系統可以互相檢視比對，一旦發生失效，系統會自行偵測到失效
3. 失效後在人不介入(接管)的條件下仍能保持安全
4. 重新恢復運作需要靠人接管
[輔助]
1. 系統單一
2. 當系統做了錯誤判斷，沒有另一套機制，所以系統壓根不知道產生錯誤
3. 系統無法自行偵測失效
4. 全仰賴在駕駛座的駕駛人注意和隨時接管
像Waymo這類自駕車
駕駛方式很笨拙，在限定區域的開放道路運行仍有一堆bug問題(卡住、逆向)
這些bug失效無法預期哪時出現，但發生當下系統
1. 能偵測到失效
2. 在人尚未接管下保持安全性
簡單講就是滿足失效安全
直到人類接管並恢復運作
Criuse則是輾死前車撞飛的人，即便車能停下但無法確保安全性，被要求需要安全駕駛
FSD會在絕大多數(99.9%)的條件下駕駛方式更接近人類
但機器學習、AI不可能訓練到完美(可能餵越多數據反而造成某些場景應對退步)
所以當今天遇到那極低的0.1%情況
系統出錯但系統並不認為失效(也沒法偵測到失效)
就得全靠駕駛人判斷是否要立即介入接管
像是ISO 26262功能安全、ISO 21448預期功能安全(SOTIF)、FIT(Failures In Time)
這些都是政府核准自駕牌照的依據，之前也都提過了
※ 引述《LimYoHwan (gosu mage)》之銘言：
: https://tinyurl.com/bvkhn237
: 簡易翻譯：
: Cruise公司表示,他們將恢復採用有安全駕駛員監督的自動駕駛模式。這是因為他們去年
: 發生的一起無人駕駛事故後,暫時停止了無人駕駛服務,並進行了全面的安全評估。
: 經過評估後,Cruise決定先恢復自動駕駛,但仍然會有安全駕駛員在車內監督。這可以在提
: 供自動駕駛功能的同時保證乘客安全。
: https://i.imgur.com/WVhDrCF.jpeg
: Cruise表示,這種監督式自動駕駛模式只是過渡階段,他們最終的目標仍然是實現完全無人
: 駕駛。此次事故顯示,即使是先進的自動駕駛技術,在實際運營中仍需要更謹慎的安全措施
: 。
: 齁，cruise配備這麼多感知設備，一堆光達，也有備援系統，且通過政府法規認證可以運
: 行無人計程車，啊 怎麼車上又放人了，那我可以認為「人」才是備援嗎？那些光達根本
: 沒屁用…..
: 自駕的瓶頸是AI問題還是光達、硬體問題？
: 車版大師表示？
: Starting this week, Cruise AVs will begin supervised autonomous driving in Pho
: enix as Cruise takes further steps towards returning to its driverless mission
: .
: For the past several weeks Cruise has been mapping and collecting road informa
: tion in Phoenix. Now, we will validate our AV’s end-to-end behaviors against
: our rigorous safety and AV performance requirements. During this phase, the Cr
: uise vehicles will drive autonomously with a safety driver present behind the
: wheel to monitor and take over if needed.
: Supervised autonomous driving is a critical validation phase prior to driverle
: ss deployment and builds on our extensive work in simulation, closed-course dr
: iving and more than 5 million driverless miles previously driven by our fleet
: to ensure safe performance on real-world roads and driving scenarios. Safety d
: rivers play an important role in testing the AV’s performance and the continu
: ous improvement of our technology. We’ll begin supervised autonomous driving
: in the city of Phoenix and will gradually expand to Scottsdale, Paradise Valle
: y, Tempe, Mesa, Gilbert and Chandler measured against predetermined safety ben
: chmarks.
: Safety is the defining principle for everything we do and continues to guide o
: ur progress towards resuming driverless operations. From comprehensive vehicle
: management before our AVs depart the garage to continuous driver monitoring a
: nd roadworthiness of the vehicle, we deploy rigorous safety procedures, protoc
: ols and performance requirements to ensure supervised autonomous operations ar
: e safe for operation on public roads.
: Our safety measures and procedures are multifaceted and can be grouped into th
: ree pillars:
: Product Safety: Cruise AVs are fully-integrated autonomous vehicles derived fr
: om the Chevy Bolt EV platform, manufactured by General Motors, and tested to v
: alidate that the vehicles meet all motor vehicle safety requirements after th
: e addition and installation of Cruise AV hardware (sensors, etc.). We have pro
: cedures in place to ensure our AVs are appropriately maintained and that new s
: oftware is thoroughly tested and validated through a measured release process
: prior to deploying to the fleet.
: Operational Safety: We have documented procedures and protocols in place to en
: sure the safety of our fleet operations for those in our vehicles and those we
: share the road with, including driver training, ongoing performance managemen
: t processes, and responsible scaling of operations.
: Enterprise Safety: Safe operations start with a safety-focused culture, and ev
: ery worker at Cruise is empowered and responsible for the safety of our produ
: ct and operations. We enable safety culture through appropriate governance, po
: licies, reporting mechanisms, and regular safety training.
: We are committed to partnering with the communities we serve, so whether human
: -driven or autonomous, we encourage any community member to report driving con
: cerns to our Community Feedback portal. When a concern is submitted, we will i
: nvestigate the claim and, as needed, take appropriate action. Our feedback for
: m: getcruise.com/support-center
: As we continue through this journey, we look forward to working with local com
: munities to jointly achieve our shared mission of making transportation safer
: for all.
--
※ 發信站: 批踢踢實業坊(ptt.cc), 來自: 114.34.222.29 (臺灣)
※ 文章網址: https://www.ptt.cc/bbs/car/M.1715682027.A.82B.html
※ 編輯: chandler0227 (114.34.222.29 臺灣), 05/14/2024 18:21:35
→ pippen2002: 結論: 不太好?05/14 18:23
推 XXXXBANG: FSD會越來越爛 其他廠牌的只會越來越進步 05/14 18:27
推 alaokj: 錢大文也是先推再看了啦~ 05/14 19:04
→ LimYoHwan: 好像不是你說呢捏 05/14 19:34
推 bmw530li: 演算法就算聰明，但Sensor只有一套時，誤判就會車禍(演05/14 20:01
→ bmw530li: 算法完全信了)，所以才一直要求要備援機制自我驗證，而 05/14 20:02
→ bmw530li: 不是很相信自己的破攝影機就輾過去(躺著的貨櫃車表示： 05/14 20:02
→ bmw530li: 其他廠商自知還不成熟，所以沒人敢明目張膽的大吹特吹 05/14 20:02
→ bmw530li: 更沒人敢躲在Lv2放消費者去幫忙測試 05/14 20:03
推 arabeske: 類神經網路到底從訓練中學習得什麼真的不好驗證。AI算05/14 20:36
→ arabeske: 圖1個frame 出錯人眼沒看到就算了，但自駕可會出人命 05/14 20:36
→ LimYoHwan: 撞貨櫃就是兩套camera+radar 這特黑有夠不專業的 而且 05/14 20:36
→ LimYoHwan: 多久以前 還是ap 不是fsd 笑死05/14 20:36
→ LimYoHwan: 你那邊還來得及all in nVDA05/14 20:38
→ LimYoHwan: 演算法要兩套 你大腦是不是也要兩個05/14 20:46
推 StarburyX: 其實無論你用哪種方案 都逃不了AI05/14 20:57
→ StarburyX: 掌握算力 就掌握AI 就掌握自駕05/14 20:58
→ StarburyX: 現在不走端到端AI 基本就是死路05/14 20:59
E2E的確是因為特斯拉才開始廣泛被一般民眾知道
但畢竟也不是特斯拉首創或獨創
E2E背後關鍵是DRL(深度強化學習)技術
例如幾年前的圍棋AlphaGo就是用DQN結合蒙地卡羅法
DRL也很早大量應用在自駕感知
(但決策和控制模塊)
即便DRL有極大發展潛力，有效解決傳統模塊分層的諸多問題
但依舊存在實際應用安全跟缺乏可解釋性等問題
跟自駕車應用安全相關的：
Mobileye提出的RSS(Responsibility-Sensitive Safety)安全模型
Nvidia提出的SFF(Safety Force Field)
→ StarburyX: Waymo cruise都一攤死水05/14 20:59
→ StarburyX: Mobile eye 沒聲音 專注在賣車廠基本方案 05/14 21:00
→ StarburyX: 對岸華為 還有一戰之力 05/14 21:00
→ StarburyX: 解決自駕 最重要的就是算力 05/14 21:01
※ 編輯: chandler0227 (114.34.222.29 臺灣), 05/14/2024 22:14:37
推 bmw530li: 自己做不好就全世界都不准做好XDDD 05/15 00:34
→ bmw530li: 737MAX的case應該可以說明 05/15 00:35
