影视风云路

智驾辅助一招变"长板" 一汽丰田bZ5为何青睐Momenta 5.0_我的网站

速度与激情6

一 | Stockholm, Oct 10 (UNI) The Royal Swedish Academy of Sciences announced Monday that three US-based economists won the 2022 Nobel Memorial Prize in Economic Sciences “for research on banks and financial crises”.

The winners are Ben Bernanke, former head of the US Federal Reserve and currently with the Brookings Institution in Washington DC, Douglas Diamond of the University of Chicago and Philip Dybvig of Washington University in St. Louis, Missouri. The winners share a prize of about $900,000.

The nobelprize.org put out a release saying:

"Their discoveries improved how society deals with financial crises

"This year’s laureates in the Economic Sciences, Ben Bernanke, Douglas Diamond and Philip Dybvig, have significantly improved our understanding of the role of banks in the economy, particularly during financial crises. An important finding in their research is why avoiding bank collapses is vital.

"Modern banking research clarifies why we have banks, how to make them less vulnerable in crises and how bank collapses exacerbate financial crises. The foundations of this research were laid by Ben Bernanke, Douglas Diamond and Philip Dybvig in the early 1980s. Their analyses have been of great practical importance in regulating financial markets and dealing with financial crises.

"For the economy to function, savings must be channelled to investments. However, there is a conflict here: savers want instant access to their money in case of unexpected outlays, while businesses and homeowners need to know they will not be forced to repay their loans prematurely. In their theory, Diamond and Dybvig show how banks offer an optimal solution to this problem. By acting as intermediaries that accept deposits from many savers, banks can allow depositors to access their money when they wish, while also offering long-term loans to borrowers.

"However, their analysis also showed how the combination of these two activities makes banks vulnerable to rumours about their imminent collapse. If a large number of savers simultaneously run to the bank to withdraw their money, the rumour may become a self-fulfilling prophecy – a bank run occurs and the bank collapses. These dangerous dynamics can be prevented through the government providing deposit insurance and acting as a lender of last resort to banks.

"Diamond demonstrated how banks perform another societally important function. As intermediaries between many savers and borrowers, banks are better suited to assessing borrowers’ creditworthiness and ensuring that loans are used for good investments.

"Ben Bernanke analysed the Great Depression of the 1930s, the worst economic crisis in modern history. Among other things, he showed how bank runs were a decisive factor in the crisis becoming so deep and prolonged. When the banks collapsed, valuable information about borrowers was lost and could not be recreated quickly. Society’s ability to channel savings to productive investments was thus severely diminished.

"The laureates’ insights have improved our ability to avoid both serious crises and expensive bailouts,” says Tore Ellingsen, Chair of the Committee for the Prize in Economic Sciences.

The Nobel Memorial Prize in Economic Sciences was not included in Alfred Nobel's will, voanews.com reported.

It was established by the Swedish central bank in 1968 — and is officially the "Sveriges Riksbank Prize in Economic Sciences in memory of Alfred Nobel" to mark the bank’s 300th anniversary. It was first awarded in 1969 and has been funded by the bank since then.

UNI ING。

[ 新鲜技术解读] 6月10日，一汽丰田的全新纯电中型SUV——bZ5正式上市，售价12.98-15.98万元。包含四款配置，原厂自带“中国血脉”，亮点主推安全智能，外带舒适大空间的bZ5明显是为走量而来。那么新车到底有哪些技术亮点？让丰田青睐的Momenta 5.0智能驾驶辅助水平到底怎样？以及Momenta到底在做什么？本期内容我们就来聊聊。

丰田标准安全设计+先进智驾辅助，15.98万即可得到。 先看新车，bZ5与之前广汽丰田的铂智3X同样基于e-TNGA架构打造，外观也采用了家族化的“锤头鲨”前脸造型，但车身尺寸更大，且全系装配比亚迪“刀片”电池，能提供550km和630km两种续航选择，参考官宣素材，整车在电池主被动安全和控制安全方面亮点较多，例如采用了可抑制电池短路的驱动用电池冷却液，通过采用“双电路板+双电机+解耦设计”实现“制动冗余”，通过EPB+ECU双芯片+备用电源，实现“驻车冗余”等等。毕竟安全可靠一直是丰田的好牌，新车还得打。

二 |

再看内装，新车全系标配9个安全气囊，并配备了15.6英寸的超薄窄边屏幕，官方称“通过使用色域广、反射率低的低温多晶硅技术，确保更清晰鲜艳的画面”，其中高配车款还装配有1.44平米的全景天幕，采用Low-E+Ag2双层镀膜玻璃，据说紫外线隔绝率能达到99.98%，红外线隔绝率99.3%。此外，Toyota Space车机系统还集成了科大讯飞语音、AI助手、生物监测、智能香氛等亮点配置。一句话，自主品牌15万能给到的功能配置，丰田这回也都满足您。

当然更重要的：一汽丰田bZ5(参数|询价)同样采用了Momenta 5.0智能驾驶辅助算法，而让国际大厂不再固守全栈自研，转而直采本土供应商的智驾辅助新方案，Momenta 5.0凭什么？智驾辅助渐成新车刚需，丰田急需迎头赶上。

三 | 首先，在中国市场纯电动车销量势头高歌猛进和智能驾驶辅助技术快速普及的双重背景下，丰田是急需迎头赶上的，而直接“拿来”更懂道路国情的中国自动驾驶科技公司方案，细节处按需要适当优化，显然更加事半功倍。

四 | 按照丰田的官宣介绍，最新的Toyota Pilot 智能驾驶辅助是由Momenta飞轮大模型5.0＋Toyota Safety Sense智行安全合力构筑，而在全面融入Momenta 5.0智能驾驶辅助能力后，系统可以对道路识别的更多元，更符合中国实际道路状况，比如窄道无划线路识别对向来车、识别行人、电动二轮车、对向货车以及高速施工路段等，同时避障策略也更智能，比如右转遇视线盲区施工、施工场景下避让两轮车、路口左转避让车辆插入、路口鬼探头等都可以安全避险。同时，车辆还可支持遥控泊车辅助、狭窄路巷的泊车智驾辅助以及最长3000米的记忆泊车功能。

五 | 其中，城市INP导航智驾辅助的变道成功率据说可达95%以上，且能完成3车道的掉头，而高速INP的上下匝道成功率也能高达98%，上车效果可谓立竿见影。当然，像预碰撞安全、变道辅助等智能规避碰撞的任务会继续交给Toyota Safety Sense智行安全套装。

全部实现以上功能，还要靠33颗感知硬件，在智行版车款上，装配了1个激光雷达、5个毫米波雷达、12个超声波雷达、4个环视摄像头、4个侧视摄像头、2个前视摄像头、1个后视摄像头、1个乘员监测摄像头、1个疲劳监测摄像头、1个驾驶员运动脱离监测（电容感应式）以及1个高精定位单元。而芯片算力可达544TOPS（是已经上市的铂智3X两倍多）。丰田想要的第一梯队，“一个飞轮两条腿”。 首先，一汽丰田bZ5融入的Momenta 5.0属于“一段式端到端”大模型架构，最早发布于2024年9月，这个时间距离特斯拉3月底推出FSD V12端到端方案不到半年的时间，而距离4月华为ADS 3.0推出端到端更近，Momenta也称其为“飞轮大模型”，目前在智己、比亚迪、腾势、埃安等诸多新车上都有搭载，加之3.0和4.0阶段的累积装车量（据说2024年约30万辆，2025年底预计能达到约100至150万辆规模），城区辅助驾驶的量产车累计行驶里程已达20亿公里，因此，无论是架构先进性还是样本量，Momenta都是妥妥的第一梯队，而这正是丰田新车想要且需要的，当然，丰田与Momenta本身就是投资共创的关系，而青睐Momenta的也远非丰田一家，像上汽、奔驰、通用，以及广汽、比亚迪、日产等诸多主机厂也都有定点合作成果，如此多的选择，可见Momenta在智驾软件的用户体验、硬件的成本控制、算法兼容能力以及长尾问题的解决，还是很有优势。

这里多说一点Momenta多次强调过的“一个飞轮两条腿”发展策略，“飞轮”自然就是“数据飞轮”这一高效开发模式，所谓“两条腿”，指的就是两条技术路线，一条是和华为ADS相似的Mpilot方案，主攻量产车L2、L3方案，一条是解决自动驾驶RoboTaxi的MSD（Momenta Self Driving）无人驾驶方案，“两条大腿”相辅相成，用量产辅助驾驶产生海量数据去训练自动驾驶算法，解决L4完全无人驾驶的长尾问题。

六 | 同时再用L4的技术赋能L2量产辅助驾驶产品力的提升，形成技术迭代与数据收集的良性循环。从短期记忆到长期学会，Momenta 5.0只是新的开始。

简要梳理Momenta的发展历程，2019年开始用Transformer模型架构（一种深度学习领域的基础架构）做预测、路径规划；2022年发布的AD 2.0，开始用AI模型做感知，多任务预测，并开始尝试决策模块的模型化，这一时间基本属于研发试验阶段；2023年上半年AD 3.0开始面向量产车，系统感知可以将时序任务整合到同一模型，规划算法转向数据驱动；而2023年下半年，AD 4.0可以将感知所有任务整合到同一模型，规划全部数据驱动处理。这一阶段的智驾辅助方案，可以理解为“两段式端到端”。

具体到现在的Momenta 5.0，它实际是将感知与规划整合到了一起，典型的一段式端到端大模型，即输入端直接来自摄像头、雷达等传感器的原始数据，而输出端就是方向盘、制动等执行路径的规划，相比过去那种感知、识别、预测、规划再到执行，冗长的决策步骤，“端到端”相当于全程交给一个“大脑”处理，主打一个“没有中间商赚差价，时间和效率自然会大大提升，而且“看路开车”也能让智能驾驶辅助开得更像“老司机”。不过它也保留了之前的DLP（Deep Learning Planning）深度学习规划模型，即通过训练神经网络模型来处理和预测复杂的数据模式，以便快速学习，很像人类的短期记忆。