优秀工程师们总是喜欢迎接挑战，而自动驾驶汽车的出现更是为这些全球最理性却又同时最富创意的头脑带来了终极考验。

在最近接受《汽车工程》采访时，摩比斯北美 (Mobis North America) 自动汽车研发总监 David Agnew 表示，“使用案例和耐力测试非常重要，但这仍不足以最终坚定制造商的信心，真正开始生产非常复杂的自动驾驶系统。汽车制造商通常在投产前需要累计5 亿到 10 亿英里的无事故行驶里程，但自动驾驶汽车哪怕花十年时间也根本无法实现这个数字。然而回顾历史，挑战从未阻碍技术前进的脚步，未来也不会。”

摩比斯的自动驾驶出行战略

摩比斯北美公司的高级研发部门成立于 2015 年底，公司的首项创新技术将是一款名为 DDREM 的安全系统，即“司机意识丧失的救援和退出机制”。这套系统的开发主要是为了达到 5 万英里无事故目标。Agnew 表示，“这种难度比全自动驾驶汽车低 4 个数量级。但我们认为这只是一个小问题，摩比斯一定可以通过创新思维解决这个问题。”

据了解，DDREM 系统预计将在 2020 到 2021 年间做好量产准备。该系统将监测驾驶员的操作状态，并在驾驶员失去对车辆的控制能力时自动启动 SAE 4 级自动驾驶系统，引导车辆进行安全停车。具体来说，驾驶员出现打瞌睡、心脏病发作或其他严重的医疗紧急情况时均会触发该系统。

目前，在公司位于密歇根州普利茅斯市的研发中心内，摩比斯共有18 名优秀工程师和机器人技术及计算机技术专家正在进行 DDREM 系统的开发。未来几个月中，摩比斯的 DDREM 团队将进一步扩大至约 25 位技术专家。

Agnew 表示，“我们的两个先进技术工程团队有一定的工作重叠，设计团队正在按照我们的性能要求进行 DDREM 系统开发，而测试团队正在努力搞清楚我们该如何进行分析和验证，并最终确保我们的设计的确可以达到性能要求。”

然而，DDREM 自动系统的结构非常复杂，对这样的系统进行测试和验证势必面临巨大的挑战。

“从测试方面来说，投产一款汽车非常简单。”Agnew指出，因为行业已经有了一套完整的量产汽车测试标准流程。他说，“你必须对车辆的安全带、变速箱、照明等各个系统进行逐一验证，但这些量产系统的测试项目和流程均已得到充分验证，你只需按照规定进行即可。然而自动驾驶汽车的情况完全不同。”

自 2004 年，美国国防部高级研究计划局 (DARPA) 拿出 100 万美元奖金举办自动驾驶技术挑战赛以来，这项技术已经取得了长足的进步。尽管当时并没有任何团队能够按照要求打造出一款可以穿越 150 英里莫哈韦沙漠的自动驾驶汽车，但这项挑战赛却为日后的自动驾驶汽车研发与测试拉开了序幕。

目前，有多项自动驾驶汽车项目已经开始进行封闭和公共道路测试。美国地区，福特汽车已经增加了公司自动驾驶测试车队中Fusion 混合动力轿车的数量，并计划在 2021 年开始生产SAE 4 级自动驾驶汽车；而通用汽车也会将公司自动驾驶测试车队中的雪佛兰 Bolt 电动车数量从 50 辆增加至 180 辆。

最近，本田公司宣布，在 2020 年前准备好具备高速公路自动驾驶功能的 SAE 3 级自动驾驶系统；并在 2025 年前实现推出 SAE 4 级自动驾驶汽车的目标；丰田协作安全研究中心 (Collaborative Safety Research Center) 启动了11 项有关自动驾驶和互联技术的研发项目。与此同时，宝马 、菲亚特克莱斯勒、英特尔 和 Mobileye 也在今年早些时候宣布，将合作开发一个针对SAE 3 级到 4/5 级自动驾驶的可拓展化平台。此外，其他汽车制造商和供应商也在进行高级自动驾驶项目的相关研发。

模拟和智力

假设一下，如果高级驾驶员辅助系统 (ADAS)，也就是人们通往自动驾驶终极目标的重要环节，仅进行了一类测试。目前，ADAS 系统的使用案例测试仅将对有限的特定情景进行验证，例如当车辆偏离当前车道时，司机将收到一条“车道偏离警告”。

Agnew 表示：“一旦引入人工智能 (AI) 或其他复杂度更高的系统，整个情况将完全不一样。”一些看似微小的变化都有可能给自动驾驶带来巨大影响。例如，一辆自动驾驶车辆的传感器系统本来应该可以监测到模拟行人的明显腿部移动，但如果这位“行人”的腿部被外套等衣物遮挡，传感器可能就没办法正常工作，这时的系统工作情况就会完全不一样了。总而言之，工程师在开发自动驾驶系统时，必须不遗余力地确保系统能够准确识别所有驾驶情境，并做出正确的响应。

“汽车行业的传统是进行案例测试，但我们已经进入自动驾驶技术的新纪元。毫无疑问，案例测试仍将是自动驾驶汽车测试的重要组成部分，但我们也需要适当调整测试方法。”Agnew 同时强调，我们必须“尽最大可能，预测自动驾驶系统的运行状况”。

模拟仿真技术可以成千上万个变量的作用下，预测系统运行结果，这也正是该技术可以在自动驾驶技术的研发和测试中发挥重要作用的主要原因。

根据Navigant Research咨询公司高级分析师 Sam Abuelsamid 的说法，“仿真环境在验证自动驾驶系统的性能和稳定性方面至关重要，特别对软件堆栈的AI 部分，比如神经网络来说。”

协助进行“极端情况”测试

与许多其他研发自动驾驶技术的公司一样，摩比斯也认为协作是必不可少的。Agnew 透露，公司将在不久的将来与定制可扩展模拟工具提供商 Metamoto 签订合同。据这家硅谷创业公司首席执行官 Chad Partridge 介绍，Metamoto 的模拟产品可以协助工程师在自动驾驶系统中完成数百万次的日常测试，并“智能地探索极端情况下的参数范围和性能极限。”

仿真软件对于解决“极端情况”的困扰至关重要。 “极端情况”是指一些几乎不可能提前准备，但又的确有可能发生的情况。当工程师在考虑对自动驾驶系统进行改变，比如修改 AI 设计或者传感器架构时，仿真环境可以帮助工程师在真正着手进行修改前，以一种非常安全的方式评估这种改变可能带来的影响。

总部位于康涅狄格州斯坦福德市的高德纳研究咨询公司 (Gartner) 研究总监Michael Ramsey 表示：“仿真软件还可以从无数个不同角度，对一项技术进行验证，而且速度比真实测试更快，成本也更低。”

移动模拟

仿真环境可以为工程师提供一个二维界面，这在判断最坏情况时尤为有效。不过，未来摩比斯的工程师和研究志愿者还将在荷兰 Cruden 公司定制的新型运动模拟器中，体验 6 维浸入式模拟驾驶环境。具体来说，Cruden 公司的模拟器共有 6 个维度，预计将于 2018 年 4 月正式在摩比斯的密歇根中心投入使用。Agnew 解释说，“我们认为，固定式模拟器工作站可能很难为我们提供最全面的人机界面 (HMI) 数据。”

通过运动模拟器内部的摄像头和传感器，工程师可以观察并收集驾驶员在开车睡着时的数据。这些信息将为一批大学研究项目的车内视频记录及其他相关研究项目提供素材。Agnew 表示：“我们希望观察到，在 DDREM 系统激活后，油门、刹车等车内组件在驾驶员睡着后的反应，及驾驶员清醒后试图重新控制车辆时的情况。

工作站所提供的模拟仿真环境，将为摩比斯的工程师提供一个研发DDREM 系统及其运行算法的框架平台，而运动仿真器生成的数据将重新用于进一步的仿真，并支持耐久性测试和其他案例测试。

摩比斯北美公司还将在一辆由公司技术人员和制造团队共同打造的自动驾驶汽车上进行大量测试。这辆展示车是摩比斯工程师在韩国制造的自动驾驶汽车复制版。Agnew 表示：“我们预计将在今年年底前，在美国引入首款用于DDREM 研发的自动驾驶原型车。”

届时，原型演示车将协助进行第一阶段测试，内容包括车辆自动变道——也就是说，车辆将根据情况进行自动变道，并在侦测到道路边缘后安全停车。Agnew 表示：“在自动驾驶汽车研发中，与功能开发相比， 要100%确定自动驾驶汽车何时可以正常行驶，何时可能出现故障，才是最难的。”

Engineers like a good challenge, and the development path for autonomous technologies is putting logical and creative thinkers to the ultimate test.

“Use case and endurance testing—although highly relevant—are not enough to get to the confidence level that’s needed before going into production with a highly-complex automated driving system. When you’re trying to hit a metric that’s 500 million to a billion miles of accident-free driving, you just can’t drive that many miles, even in a decade. Challenges haven’t stopped engineering progress in the past, and it isn’t stopping progress now,” David Agnew, Director Advanced Engineering, Autonomous Vehicle Research for Mobis North America, asserted in a recent interview with Automotive Engineering.

Automated mobility for Mobis

The first novel autonomy technology from Mobis North America’s advanced engineering group, formed in late 2015, will be a safety-specific system known as the Departed Driver Rescue and Exit Maneuver (DDREM). The system is being developed to a one-in-50,000-mile accident-free metric: “That’s four orders of magnitude easier than doing a full autonomous-vehicle metric. We think that it’s a small enough problem that Mobis and out-of-the-box thinking can solve,” said Agnew.

A production-ready DDREM system is expected in the 2020-2021 timeframe. The system will detect when a driver is no longer operating the vehicle, then initiate SAE Level 4 autonomous control to guide the car to a safe stop. A driver who has fallen asleep will trigger system activation; engineers are also considering other system activation triggers, such as a heart attack or other severe medical emergency.

Currently, 18 Mobis engineers, robotic, and computer science specialists working at the firm’s Plymouth, Michigan, facility are tasked with developing DDREM. In the coming months, the team will be expanded to approximately 25 technology experts.

“There is a definite overlap of work between our two advanced engineering teams. The design team is working to develop DDREM to our performance requirements, while the testing team is working to figure out how we’re going to analyze, validate, and build confidence that the design will meet the performance requirements,” said Agnew.

DDREM’s development effectively underscores the challenges associated with testing and validating a highly-complex autonomous system.

“To put a car into production right now, it’s pretty straightforward in terms of testing,” said Agnew, noting that use case physical testing is standard practice across the industry. “You verify product designs for seatbelts, transmissions, lighting and other systems with specific tests. But all of the tests for those in-production systems have already been verified. That’s just not the case with autonomous technologies.”

Progress in developing self-driving vehicles has been substantial since the Defense Advanced Research Projects Agency (DARPA) first offered a $1 million prize in its 2004 challenge. Although no team created an autonomous vehicle capable of completing the required 150-mile trek through the Mojave Desert, that event was a touchstone for focused autonomous-system development and testing.

The scorecard of current development projects includes many programs with closed-circuit and public-road testing. In the U.S., Ford is upping the number of Fusion Hybrid sedans in its autonomous test fleet and plans to start producing SAE Level 4 self-driving vehicles in 2021, while General Motors will increase its autonomous-driving test fleet of Chevrolet Bolt electric cars from 50 to 180.

Honda recently announced plans for highly-automated, SAE Level 3 expressway-driving capability by 2020—with SAE Level 4 capability targeted for 2025. Toyota’s Collaborative Safety Research Center recently launched 11 research projects for autonomous and connected-vehicle technologies. Meanwhile, BMW, FCA, Intel and machine-vision specialist Mobileye earlier this year announced a cooperative program to develop a scaleable SAE Level 3 to Level 4/5 platform. Other automakers and suppliers are also in advanced autonomous-development programs.

Simulation and brainpower

Consider just one type of testing conducted on advanced driver-assistance systems (ADAS), the forerunner to integrating these technologies for autonomous operation. Current use case tests for ADAS verify certain actions, such as the driver receiving a lane-departure warning when the vehicle departs from the current traffic lane.

“With artificial intelligence (AI) or systems with a high degree of complexity, it’s a completely different game,” said Agnew. Autonomous-driving scenarios can be impacted by seemingly small variations. For instance, an autonomously-driven vehicle’s sensor system can detect visibly moving legs on a pedestrian dummy. But the scenario changes if the pedestrian dummy’s legs are covered by an overcoat and that clothing confounds sensor detection. The bottom line is the developers of an autonomous systems must strive to design a system capable of recognizing and correctly responding to all possible driving scenarios.

“The auto industry works in a use-case environment, but we’re in new territory with autonomous technologies. While use-case testing still has a vital role, we also need to adapt how we test,” said Agnew, stressing the need “to predict as much as we can about how well the autonomous system is going to work.”

An ability to predict outcomes amid hundreds of thousands of variables is the primary reason why simulation environments are emerging to assume major role in autonomous technology development and testing.

“Simulation environments will be absolutely crucial in validating the performance and consistency of automated driving systems, especially the AI components of the software stack, such as the neural nets,” according to Sam Abuelsamid, Senior Analyst-Energy for Navigant Research in Detroit.

Help with 'corner cases'

Like many other companies developing automated-driving technology, Mobis views collaboration as an essential. In the near-term, Agnew expects to ink a contract with Metamoto, a provider of purpose-built scaleable simulation. Metamoto’s simulation offering will enable engineers to conduct millions of daily tests on autonomous systems, “while intelligently exploring parameter spaces and performance boundaries across relevant edge cases,” according to Chad Partridge, CEO of the Silicon Valley startup firm.

Simulation software is vital for ironing-out the "corner cases"—situations that are nearly impossible to plan for but could happen. If engineers want to alter the autonomous system, such as the AI design or the sensor architecture, the simulation environment is a non-risk way to evaluate the effect of a change prior to targeted physical testing.

“Simulation software also can be used to validate numerous different aspects of the technology much faster and less expensively than real-world testing,” according to Michael Ramsey, Research Director of Gartner, Inc., a Stamford, Connecticut-headquartered research and advisory company.

Simulation on the move

Simulation environments provide a two-dimensional interface and are especially helpful in determining worse-case scenarios, but Mobis engineers and research volunteers will experience an immersive environment of replicated roadways and vehicle movements inside a new motion simulator being custom-built by Cruden of the Netherlands. The 6-degrees-of-freedom motion simulator is expected to be operational at Mobis’ Michigan facility in April 2018. “We didn’t feel that our static simulation workstations could provide us with comprehensive human machine interface (HMI) data,” Agnew explained.

Cameras and sensors inside the motion simulator will enable engineers to observe and collect data of a driver falling asleep at the wheel. This information will compliment in-vehicle video recordings from university research programs as well as other studies on drowsy drivers. “We’re interested in what happens to the accelerator pedal and other in-vehicle controls while the driver is asleep as well as what can happen if the driver wakes up and attempts to regain control of the vehicle after DDREM activation,” said Agnew.

While workstation simulation environments will provide a framework for Mobis engineers to develop the DDREM system and its operational algorithms, the motion simulator will generate data to be looped back to simulation as well as provide information that will be applied to endurance and use-case tests.

Mobis North America also will conduct tests on a demonstrator autonomous vehicle built in-house by Mobis technicians and fabricators. This vehicle is a duplicate of a prototype autonomous vehicle built by Mobis specialists in South Korea. “We expect to have our first prototype autonomous vehicle in the U.S. for DDREM development by the end of this year [2017],” said Agnew.

Phase-one track testing with the prototype demonstrator will involve autonomous lane shifts until the car detects the edge of the roadway and performs a safe stop. “Developing the autonomous capability will be easy in comparison to understanding fully how often the self-driving maneuvers will work and how often those maneuvers will fail,” Agnew said.

Author: Kami Buchholz
Source: SAE Automotive Engineering Magazine