Autonomous Driving Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
Autonomous Driving Moonshot Project with Quantum Leap from Hardware to Software & AI Focus 02 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
Summary 04 Autonomous Driving: Hype or Reality? 06 Deep dive: Artificial Intelligence 18 Impact on Today’s Automotive Industry 28 1. Product Structure 35 2. Work Structure follows Product Structure 38 3. New Steering Models 47 4. Mastering New Technologies 50 The Way Forward 54 Authors 56 Endnotes 57
03 Summary
Future autonomous (electric) vehicles are primarily software-driven products compared to traditional cars. The upcoming transformation in the automotive indus- try from a “made of steel” business towards “software is eating the world” will be no doubt a game changer – for better or worse. Now that new players from the tech sector have entered the stage in the automotive industry, traditional manufacturers and suppliers try hard to continuously shorten development cycles and to catch up with the inevitable move into the new software era. Collaborative agile working models predominantly known from the software industry and more innovative cooperation management approaches are paving the way for tackling these challenges and turn them into opportunities.
04 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
05 Autonomous Driving: Hype or Reality?
In recent years, autonomous driving and so-called robo- taxis have become one of the hottest topics in the auto- motive industry - and beyond! Traditional car manufactur- ers and established suppliers are not the only ones who are trying hard to find the sweet spots in this new emerg- ing mobility value chain.
Tech giants like Nvidia and Intel, leading increase in e-mobility, which the world is still software and internet players like Google waiting for, and lately blockchain and Bitcoin, (Waymo) and new mobility startups such which receive significant media attention. as Aurora, Cruise and Uber are also on But where among all these trends and hypes the verge of reaping the rewards of an can we place autonomous driving? The entirely new future mobility era. Unlike the following paragraphs show some forecasts stakeholders of today's automotive industry, to frame the general market potential for they do not have vested stakes to protect. autonomous driving solutions and provide However, on the other hand we are all aware a framework to align on common terms and of several technological hype cycles, ranging wording when it comes to automated and from the internet bubble at the turn of 'real' autonomous driving. the millennium, the proclaimed significant
06 07 Voices on Autonomous Driving enthusiasm. However, there has also been to massively change the way we live and the Autonomous driving is receiving significant some bad press, mostly because of fatalities significance it has owing to the fact that we media attention, not least because traditional due to technological errors (see Figure 1). as humans give away control and thus put car manufacturers and tech giants are invest- Some argue that these are individual our lives in the hands of an algorithm. It will ing heavily in new technologies and prom- cases and should not distort the fact that need time and positive reinforcement for ising start-ups or forging new partnerships, statistically speaking, autonomous driving the general public to ultimately accept and but also because of significant technological is already safer than normal driving. At the trust this new technology – advances. Overall, public perception is same time, autonomous driving is under the and its inventors. positive and surrounded by an optimistic scrutiny of the public eye due to its potential
Figure 1 – Voices on autonomous driving
Where are we today?
»Exclusive: BMW to introduce »Automated vehicles may bring a new »Volvo and Baidu join forces to ‘safe’ fully autonomous driving breed of distracted drivers«2 mass produce self-driving electric by 2021 with iNext«1 ABC News, 24 Sep. 2018 cars in China«3 Digital Trends, 28 Sep. 2018 CNBC, 1 Nov. 2018
»Why People Keep Rear-Ending »Tencent builds autonomous driving »Have autonomous vehicles Self-Driving Cars«6 team in Silicon Valley«4 hit a roadblock?«5 Wired, 18 Oct. 2018 Financial Times, 7 Nov. 2018 Forbes, 1 Nov. 2018
»Amazon Could Become the Next Big »Self-Driving Cars Can Handle Player in Autonomous Driving«7 Neither Rain nor Sleet nor Snow«8 The Street, 24 Sep. 2018 Bloomberg, 17 Sep. 2018
»Uber asks for permission to restart self- driving car tests in Pittsburgh eight months »Waymo Shifts To 'Industrializing' Self-Driving 10 after its test vehicle killed an Arizona pedestrian«9 Tech as Robotaxi Launch Nears« Daily Mail, 2 Nov. 2018 Forbes, 6 Sep. 2018
The automotive industry is rapidly moving forward and undergoing massive change – automotive companies, tech giants, start-ups and others are working hard on solutions
Positive developments Controversial developments
08 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
The Road Towards 'Real' Future state 3 embraces the driverless Autonomous Driving revolution. Autonomous driving technology Based on Deloitte's Future of Mobility study, proves to be viable, safe, convenient and we envision four different personal mobility economical, yet private ownership continues futures emerging from the intersection of to prevail. two critical trends: Vehicle control (driver vs. autonomous) and vehicle ownership (private Lastly, future state 4 envisions a new age of vs. shared), as depicted in Figure 2. accessible autonomy. A convergence of both autonomous and vehicle-sharing trends Our analysis concludes that change will will likely lead to new offerings of passenger happen unevenly around the world, with dif- experiences at differentiated price points. ferent types of customers requiring different The earliest adopters seem likely to be ur- modes of transportation. So all four future ban commuters, using fleets of autonomous states of mobility may well exist simultane- shared vehicles combined with smart infra- ously. Future state 1 describes the status structures to reduce travel time and costs. quo in many markets, where traditional personal car ownership and driver-driven vehicles are the prevailing norm. While incorporating driver-assist technologies, this vision assumes that fully autonomous driving will not become widely available anytime soon.
Future state 2 anticipates continued growth of car and ride sharing. In this state, economic scale and increased competition drive the expansion of shared vehicle servic- es into new geographic territories and more specialized customer segments. The costs per mile decrease and certain customer segments view car and ridesharing as more economical and sustainable for getting around, particularly for short point-to-point movements.
09 Figure 2 – Autonomous driving is the main driver of future mobility
Future states of mobility
Autonomous 3 4
Personally Shared Owned Autonomous Autonomous ~$0.46/mileI ~$0.31/mileI
Low Asset efficiency High Vehicle control
1 2
Personally Shared Owned Driver-Driven Driver-Driven ~$0.97/mileI ~$0.63/mileI
Vehicle ownership Driver
Personal Shared
I~$X.XX cost estimate per mile based on US market example IIADAS: Advanced Driver Assistance Systems Source: Deloitte FoM study 2018; Statista; IHS Automotive
10 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
Currently, broad acceptance of autonomous vehicles seems much further away than a wide adoption of car and ridesharing. ADASII system Sources of potential delay include the need Bn. $ to address existing technological limitations, 30 such as proper functioning of sensors in 26 all weather conditions and comprehensive availability of high definition maps, as well 20 18 as concerns over cyber security and liability. 13 On the other hand, ridesharing services in 10 8 particular have a strong economic incentive to accelerate the adoption of autonomous vehicles, since it could significantly reduce 0 2016 2019 2022 2025 one of the biggest operational cost in their system: the driver! This is one of the main reasons why tech players like Google do not Autonomous Vehicle Sales rely on step-by-step driver-assist progres- sion as most industry forecasts predict 33 Annual # of Sales in Mn. (Figure 2: e.g. tripling of advanced driver 30 assist system (ADAS) revenues between 2016 and 2025), but instead try to jump 20 immediately to fully autonomous driving. Rather than following the historical pattern of technological innovation, autonomous 10 driving could constitute a step-change in de- 1 velopment. However, the majority of indus- 0 try experts expect the inflection point for 2021 2025 2040 widespread adoption of 'real' autonomous vehicles not before 2030. The following paragraph explains what we mean by 'real' autonomous driving.
11 Classification of Autonomous Figure 3 – Vehicle automation levels Driving Levels In terms of enabling technologies, automat- Partial Conditional No Automation Driver Assistance High Automation Full Automation ed driving is an evolution from the advanced Automation Automation driver assistance systems (ADAS) for active safety, which have been developed over Level 0 Level 1 Level 2 Level 3 Level 4 Level 5 recent decades and are still being contin- No system “Feet-off” “Hands-off” “Eyes-off” “Brain-off” No driver uously improved. A classification system based on six different levels, ranging from fully manual to fully automated systems, was
Z published in 2014 by SAE International, an Z Z automotive standardization body (compare Figure 3). Level zero to level two requires a human driver to monitor the driving environment at all times. Level zero means no driver assistance at all, while level one provides simple support like speed control.
Level two combines lateral and longitudinal Driver needs to be Driverless during control by the vehicle in specific situations. Driver is in moni- ready to take over as defined use cases Vehicle in charge of toring mode at all However, the driver needs to monitor the Driver in charge of a backup system lateral and times car and traffic at all times and be ready to longitudinal or longitudinal control Vehicle in charge of take over vehicle control immediately. Driver completely in lateral control Vehicle in charge of in all situations. lateral and charge lateral and According to use longitudinal control Vehicle in charge of longitudinal control case, no steering in many situations. lateral and in many situations. wheel and/or pedal Vehicle takes over Minimized risk for longitudinal control Warns driver in a required. other tasks accidents. in specific situations timely manner.
<2000 2000+ 2010+ 2018+ 2019+ 2021+
Source: Deloitte research 2018, SAE International 2014
12 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
The focus of current developments by car Roadmap to manufacturers is in the range from level 2 'real' Autonomous Driving to level 4. The most important transition is Automotive manufacturers are forging the Partial Conditional No Automation Driver Assistance High Automation Full Automation Automation Automation between partial automation (level 2) and path to high and full automation based on conditional automation (level 3), since in the previous experience regarding driver assis- Level 0 Level 1 Level 2 Level 3 Level 4 Level 5 latter case the driver is allowed to be out tance systems, where automation at level No system “Feet-off” “Hands-off” “Eyes-off” “Brain-off” No driver of the loop. The main difference between 2 has been realized successfully. However, level 4 (high automation) and level 5 (full the quantum leap in system reliability is automation) is the system's capability to between level 3 and level 4. At both levels, handle specific restricted driving modes vs. the system is already in charge of monitor-
Z Z all driving modes (eventually, these types of ing the driving environment, but at level 3 Z vehicles will not have a steering wheel at all). (conditional automation), a human driver still needs to be prepared to take control of the vehicle within a couple of seconds. At level 4, the system must be able to manage specified traffic conditions without any driv- er intervention and to reach a safety fallback
Driver needs to be Driverless during state in the case of unexpected events. Driver is in moni- ready to take over as defined use cases Vehicle in charge of toring mode at all Driver in charge of a backup system lateral and Figure 4 shows a roadmap towards 'real' times longitudinal or longitudinal control autonomous driving (level 5) for passenger Vehicle in charge of Driver completely in lateral control Vehicle in charge of in all situations. cars, expected to hit the road with wide- lateral and charge lateral and According to use spread adoption not before 2030. On the longitudinal control Vehicle in charge of longitudinal control case, no steering in many situations. other hand, level 3 and level 4 market intro- lateral and in many situations. wheel and/or pedal Vehicle takes over Minimized risk for duction are already underway, with level 3 longitudinal control Warns driver in a required. other tasks accidents. being primarily focused on an automated in specific situations timely manner. highway pilot and level 4 on specified appli- cations such as automated valet parking or first robo taxi fleets in selected cities (e.g. <2000 2000+ 2010+ 2018+ 2019+ 2021+ Phoenix, operated by Waymo).
13 Figure 4 – Roadmap towards 'real' autonomous driving
... 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 ...
“Google sibling Waymo Autonomous private Level 5 launches fully autonomous vehicles on public roads ride-hailing service”11
Urban and suburban pilot
Highway autopilot including Level 4 highway convoy Automated valet parking
Parking garage pilot
Traffic jam GM 201919/ Level 3 Highway chauffeur chauffeur Bosch Audi 202020/ AVP 201815 Daimler 202021/ e.Go (ZF) BMW 202122 Mover 2019+16/ Traffic jam Level 2 Conti CUbE assist, … Audi AI 17, 18 Traffic Jam Pilot 2019+ 2019 13, 14
Tesla Autopilot ACC, Stop Level 1 v9 201812 & Go, …
Level 0
Testing and ramp-up phase Source: ERTRAC 2017 “Automated Driving Roadmap”, VDA 2018 “Automatisiertes Fahren”, Deloitte Research 2018 Series deployment
14 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
The market introduction and adoption rate Key Challenges of level 3 systems and above will differ by The progression from level 3 to level 4 is market and region, because not only do cer- not a steady one. Classic rule-based ADAS tain technological issues need an ultimate functions reach their limits with level 3 solution (e.g. how to cope with extreme requirements. Linear "if then" conditions weather conditions like snow, heavy rain need to consider every possible use case or fog etc.), but regulatory and customer or combination of use cases in any given acceptance issues must also be addressed traffic situation, which is virtually impossible sufficiently upfront. in urban environments (level 4 and 5). Apart from confined spaces such as highways, traf- fic situations are highly dynamic and com- plex. For this reason, self-learning systems based on artificial intelligence (AI) that mimic human decision-making processes are critical for meeting the demand for complex scene interpretation, behavior prediction and trajectory planning. AI is becoming a key technology in all areas along the auto- motive value chain and is paramount for the success of level 4+ AD systems. Figure 5 illustrates the leap forward in technological progression from traditional software devel- opment to artificial intelligence.
However, AI talent is scarce and the market is highly competitive, resulting in skills shortages. This puts additional pressure on automotive companies still needing to build up expertise in those areas. But what exactly is AI in the first place? We are going to shed some light on this question in the next chapter.
15 Figure 5 – Quantum leap from classic rule-based coding to artificial intelligence
Maneuver 1
Maneuver 2
Maneuver Autonomous … Driving Level 3+ (Machine Learning)
Maneuver n Technology Performance Level
Substitution of rule-based coding with training of Artificial Intelligence algorithms
Classical ADAS functions (rule-based coding)
2018 2025+ (Today)
Source: Deloitte Research 2018
16 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
Figure 5 – Quantum leap from classic rule-based coding to artificial intelligence
AI is crucial for AD level 3+, because classic coding is not sufficient to master and meet the necessary requirements
•• Traffic situations arehighly dynamic and complex •• Sensors collect tremendous amounts of data points, which »Artificial Intelligence is the need to be interpreted in real time (e.g. object detection, 23 reduced visibility, natural behavior, map adjustments, …) new electricity.« •• The required processing speed for the amount of complex, Andrew Ng (AI Entrepreneur, Adjunct Professor Stanford University) new data input can only realistically be mastered with self-learning systems
•• Deep learning systems can be trained to mimic human decision-making processes, which could ease human- machine interaction on the road
•• Behavior prediction of other road users including vehicles, pedestrians, cyclists
•• AI is becoming a key technology in all areas along the automotive value chain
17 Deep dive Artificial Intelligence
Artificial intelligence is at the top of its hype curve. With potential- ly revolutionizing applications in almost every industry and do- main, the market is experiencing explosive interest from estab- lished companies, research institutions and startups alike. The global automotive artificial intelligence market forecast shown in Figure 6 reflects this interest with a CAGR of 48% between 2017 and 2025, culminating in a total volume of around 27 billion U.S. Dollar in 2025.
18 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
Figure 6 – Global automotive AI market forecast
30 27
25
21 Service 20 +48% 16 CAGR 15 12 Software
10 8 5 5 3
Automotive AI Market Size [billion USD] 2 Hardware 1 0 2017 2018 2019 2020 2021 2022 2023 2024 2025
Historic data Forecast
Source: Tractica 2018, Deloitte Research 2018
At the same time, there could not be a model based on machine learning, which artificial intelligence realm. As Elon Musk greater gap in experts' opinions on the has been around for years and would usual- put it: "The pace of progress in artificial technological short-term potential of artifi- ly classify as data mining, is now rebranded intelligence (I’m not referring to narrow AI) cial intelligence, which ranges from simple as 'AI'. Companies follow such strategies is incredibly fast. Unless you have direct performance improvements in today's to tap into the hyped sales potential. One exposure to groups like Deepmind, you have methods to artificial intelligence-powered of the prevailing reasons is that the term no idea how fast - it is growing at a pace robots conquering and enslaving the human artificial intelligence is ill-defined. There is no close to exponential. The risk of something race one day. single, agreed-upon definition that removes seriously dangerous happening is in the five- all doubt; rather all definitions leave room year timeframe. 10 years at most.” While there are promising advances across for interpretation and therefore room the entire spectrum, we see an inflation of for deceptive product specifications. This technologies branded as artificial intelli- should by no means diminish the impressive gence. For example, a demand prediction advances and speed of development in the
19 In order to separate hype from reality, we ligence. Common to most definitions is that will classify artificial intelligence in the broad- "intelligence" refers to the ability to sense er context of science, which includes but is and build a perception of knowledge, to not limited to computer science, psychology, plan, reason and learn and to communicate linguistics and philosophy. Figure 7 shows in natural language. In this context, it also the key characteristics of an AI system. The comprises the ability to process massive common understanding of artificial intelli- amounts of data, either as a means of train- gence is that it is used to get computers to ing AI algorithms or to make sense of hidden do tasks that normally require human intel- information.
Figure 7 – Key characteristics of an AI system
Big Data Reasoning
Capable of processing Ability to reason (deductive massive amounts of or inductive) and to draw structured and unstruc- inferences based on the tered data, which may situation. Context-driven change constantly. awareness of the system.
Learning Problem-solving
Ability to learn based on Capable of analyzing and historical patterns, expert solving complex problems input and feedback loops. in special purpose and general purpose domains.
Source: Deloitte 2018 “Artificial Intelligence”
20 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
We differentiate between narrow and gen- Figure 8 – 'Machine learning', 'Methods' and 'Technologies & Infrastructure' eral artificial intelligence. Today's artificial in the context of AI intelligence solutions are almost exclusively narrow. In this context, narrow means that an AI algorithm only works in the specific Artificial Intelligence context it was designed for, e.g. computer Ability to sense, reason, vision-based object detection algorithms engage and learn in autonomous driving systems. Such algo- Voice recognition Computer vision rithms have the potential to exceed human performance by orders of magnitude. Planning & Natural language General AI, on the other hand, refers to the optimization processing more human interpretation of intelligence in Machine Learning the sense that such AI solutions are able to Ability to learn understand, interpret, reason, act and learn Robotics & Knowledge Unsupervised learning from any given problem set. An AI system motion caputre typically combines machine learning and other types of data analytics methods to Reinforcement Supervised achieve AI capabilities (Figure 8). learning learning Methods Ability to reason Regression, decision trees etc.
Technologies & Infrastructure Physical enablement
Platform, UX, APIs, sensors etc.
Source: Deloitte Research 2018
21 Figure 9 – Autonomous vehicle disengagement report statisticsI
20,000 Period: December 1, 2017 to November 30, 2018
16,000
12,000
[values in km] 8,000
4,000 Autonomous km driven per disengagement
0 Waymo GM Zoox Nuro Pony.AI Nissan Baidu AImotive AutoX WeRide Cruise USA Technologies
Source: “Autonomous Vehicle Disengagement Reports 2018”, DMV, CA
I Please note, that this figure is only indicative, given that it only shows the state of California. Some firms, including German OEMs, 22 do not test their vehicles in California and therefore do not appear in the data Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
•• Based on the 2018 “Autonomous Vehicle Artificial intelligence is one of the crucial Technological Hurdles Disengagement Reports” published by the elements for level 4 and 5 autonomy. Recent On-board & Off-board DMV, California, Waymo leads the race for autonomous vehicle disengagement reports Technological hurdles for level 3 automation autonomous driving leadership by far issued by the Department of Motor Vehicles, and above are still manifold. We differentiate California, illustrate the autonomous miles between on-board and off-board challeng- •• GM’s investments in Cruise Automation driven before disengagement becomes nec- es, as shown in Figure 10. As far as on-board helped them propel to second place in essary, in critical or non-critical situations issues are concerned, the key challenges terms of autonomous kilometers driven (Figure 9). While many factors come into play revolve around sensors, computing hard- per disengagement here, it is undeniable that the firms known ware, basic software and autonomous to be strong in AI are leading the statistics. driving core software. In order to ensure the •• Zoox following in third place with a Please note that the DMV only registers safety requirements imposed by industry wider gap firms that perform test drives in the state and government, sensor quality still needs of California. The graph therefore does not to be improved to cater for e.g. accuracy for •• Noticeable advancements especially from represent the full picture, but rather serves speeds up to 130 km/h and in some cases, startups and tech companies illustrative purposes. Based on our experi- especially with lidar, the price point is still ence, the relation between top performers too high to be economically feasible. With •• Audi, BMW, Volkswagen, Tesla, Ford not and mid to low performers is accurate new central processing units and operating considered in this chart due to lack of test though. systems come new challenges regarding the data tracked by the DMV vehicle's overall safety concept. ECUs need to process large amounts of input data, but also compute complex algorithms based on artificial intelligence techniques, such as convolutional neural networks (CNN) for object detection, in real time. Considering the industry-wide trend towards electric drivetrains, the ECU's requirement for computing power is counteracted by the de- mand for low energy consumption. In terms of software development, the challenge lies in creating, training and securing (validation and verification) safe algorithms.
23 Figure 10 – On-board and off-board challenges in autonomous driving On-board hardware On-board software components
Camera/Optics System on a Chip (SOC) High Definition On-Board Maps Adaptive Cruise Control Emergency Braking Collect optical images to be inter- High performance energy- Precise localization information about Pedestrian Detection preted by advanced AI & analytics efficient computer hardware roads, infrastructure and environment Collision Avoidance
Radar V2V/V2I Localization & Mapping Determines speed and distance of Communication with vehicles & Data fusion for vehicle localization and Traffic Sign Recognition Lane Departue Warning objects using electromagnetic waves infrastructure over short range environment analysis
LiDAR Actuators Perception & Object Analysis High resolution sensor using Translation of electronic signals Algorithms Detection and classi- Cross Traffic Alert light beams to estimate distance into mechanical actions fication of objects and obstacles from obstacles Park Assist Ultrasonic sensors GPS Prediction Foresight of movements Surround View Surround View Short distance object Localization of vehicle and actions by vehicles, pedestrians and recognition (e.g. parking) using satellite triangulation other moving objects
Odometry Sensors Measure wheel speed to predict vehicle Decision-Making Blind Spot Detection travel and complement localization Planning of vehicle route, maneuvers, acceleration, steering and braking Park Assist Rear Collision Warning Off-board hardware / software Vehicle Operating Systems Data Center AV Cloud Operation Operating system running Storage and processing of Learning, adopting and up- algorithms in real time hot and cold vehicle data dating HD maps & algorithms
Park Assistance/Surround View Supervision platform
Source: Deloitte 2018 Analytics to monitor the AV system operation, detecting & correcting faults 24 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
On-board hardware On-board software components
Camera/Optics System on a Chip (SOC) High Definition On-Board Maps Collect optical images to be inter- High performance energy- Precise localization information about preted by advanced AI & analytics efficient computer hardware roads, infrastructure and environment
Radar V2V/V2I Localization & Mapping Determines speed and distance of Communication with vehicles & Data fusion for vehicle localization and objects using electromagnetic waves infrastructure over short range environment analysis
LiDAR Actuators Perception & Object Analysis High resolution sensor using Translation of electronic signals Algorithms Detection and classi- light beams to estimate distance into mechanical actions fication of objects and obstacles from obstacles
Ultrasonic sensors GPS Prediction Foresight of movements Short distance object Localization of vehicle and actions by vehicles, pedestrians and recognition (e.g. parking) using satellite triangulation other moving objects
Odometry Sensors Measure wheel speed to predict vehicle Decision-Making travel and complement localization Planning of vehicle route, maneuvers, acceleration, steering and braking
Off-board hardware / software Vehicle Operating Systems Data Center AV Cloud Operation Operating system running Storage and processing of Learning, adopting and up- algorithms in real time hot and cold vehicle data dating HD maps & algorithms
Supervision platform Analytics to monitor the AV system operation, detecting & correcting faults
25 While some companies see level 3 func- be a swift approach to achieve this amount tionalities as an evolution of classic ADAS of mileage in real world testing. Simulations functions, which can be mastered with effectively contribute to this requirement, rule-based coding, level 4 and 5 autonomy covering more than 95% of the mileage require artificial intelligence to cope with demand. However, it remains a challenge to the complexity of traffic situations. The set up the proper test concept and collect or latter typically demand large data sets (e.g. create sufficient data for validation. raw sensor data) for training, testing and validation of (deep learning) algorithms. In Overall, the challenges for companies work- order to store and process these data, com- ing on autonomous driving technology are panies make use of data centers or cloud significant and vastly affect the dynamics of solutions. The data are labelled, clustered the automotive industry. The following par- and ultimately used to optimize and update agraph discusses our view on some of the algorithms. It remains an open challenge to- key implications confronting the automotive day to efficiently validate artificial intelligence industry. algorithms such as CNNs. AI algorithms operate like a black box in the sense that it is not trivial to determine what triggers certain decisions. Validating correct functionality is cumbersome and today only feasible statis- tically via numerous test cases.
Besides, data centers constitute the basis for simulation purposes. Theoretical estimates show that in order for level 3+ autonomous vehicles to achieve approval for commercial use, the system needs to undergo billions of kilometers of testing. It is neither economical nor does it prove to
26 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
»We're entering a new world in which data may be more important than software.«24 Tim O’Reilly (Founder O'Reilly Media)
27 Impact on Today’s Automotive Industry
Cars are no longer merely the means of getting from point A to point B, nor are they simply status symbols; instead, they have become functional assets. Particularly in recent years, car manufacturers have discovered growing customer demand for digital infotainment solu- tions and other in-car services. In the telecommunications industry, smartphones replaced the traditional mobile phone, with telephony being just one of many features and oftentimes not even the most important one.
Car manufacturers face a similar trend nowadays, namely that the car is turning into a platform to serve a variety of functions. As Nitesh Bansal, Senior Vice President and Head of Manufacturing Practice Americas and Europe at Infosys, put it: “The modern car is a supercomputer on wheels, and its sensors and cameras generate a wealth of data that someday might be worth more than the automobile itself”25 – in that, for car manufacturers, the car is becoming a rich source of data they can use to improve products and business operations.
28 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
29 The same goes for development: 50 years Figure 11 – From hardware to software focus ago, the distribution of a car's added value between hardware and electrics, electronics Average automotive product cycle time (E/E) and software was approximately 95% compared to 5% respectively. In an average car today, the distribution is closer to 50% hardware and 50% E/E and software. Along with the technological progression of the 8 years semiconductor industry, development of E/E as well as software has increased exponen- tially over the past two decades. At the same 7 years time, the average product cycle time has halved over the same period (see Figure 11). 6 years
5 years
4 years
? Average automotive product cycle time
2000 2005 2010 2015 2020 2025 Year Source: Diez (2015), Deloitte Research 2018
30 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
E/E & SW share of total value add in automotive
ABS Electronic fuel On-board Hybrid Connected Autonomous injection Diagnostics cars cars cars
50%
30%
20% 15% 10% 5%
1970 1980 1990 2000 2010 2020 Year Source: Statista 2018, GTAI 2016, Brandt 2016, freescale semiconductor 2010, Wallentowitz et al. 2009, Deloitte Research 2018
That said, OEMs increasingly deviate from a around 60 control units to manage the signals and the bandwidth of bus connec- "One Product, One Function" strategy and multitude of functionalities in the vehicle, tions. Car manufacturers have started to instead approach a "One Product, Many the trend is going clearly towards a central accept the challenge and subject themselves Functions" philosophy, similar to what we processing unit that controls all functions of to - in some cases drastic - transformation have seen in the telecommunications indus- the vehicle in unison. One of the crucial chal- programs, which we will dive into in the try with the introduction of smartphones. lenges associated with a central processing following paragraph. While an average car today still features unit lies in managing the criticality of control
31 Paradigm Shift in OEM Product Figure 12 – Four major areas of organizational change Development Organizations Historically, car manufacturers have Product Structure: Work Structure follows New Steering Models: Mastering New Technologies: established a very strong top-down chain Hardware vs. Software Share Product Structure: KPI vs. Progress Indicator Cooperation vs. Supplier of command. This made sense when labor Agile Development Organization Relationship division between "thinkers" and "doers" was Processes strict. The engineer defines how the me- Analog Cockpit Silos & Waterfall KPI Supplier Steering chanic needs to assemble the car, the senior engineer instructs the junior engineer, etc. In today's VUCA world (volatile, uncertain, complex, ambiguous) these rules do not ap- ply anymore. The environment has changed, new competitors have entered the market and are constantly challenging and changing the rules and dynamics of the game. Core competencies, skills and know-how that have been perfected for decades to build great quality cars fade into the background, Digital Cockpit Cross-Functional & Agile Progress Indicator Partnership while the focus is placed on innovation, agil- ity and software, including but not limited to autonomous driving, artificial intelligence, agile working, electric vehicles and new business models.
Based on our experience, Deloitte sees four major areas of change that will likely change the dynamics of the automotive industry for good (Figure 12). •• Value add of software in the vehicle con- •• Better consideration of 'real' customer tinues to increase needs & requirements by applying mini- mal viable product (MVP) approach •• Trend is supported by surging importance of artificial intelligence (AI) •• Reduced development time, shorter time- to-market and lower development cost •• Shift away from “one function, one device” philosophy towards a “many functions, •• Ability to cope with uncertainty and com- one software platform” approach plexity
32 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
Product Structure: Work Structure follows New Steering Models: Mastering New Technologies: Hardware vs. Software Share Product Structure: KPI vs. Progress Indicator Cooperation vs. Supplier Agile Development Organization Relationship Processes Analog Cockpit Silos & Waterfall KPI Supplier Steering
Digital Cockpit Cross-Functional & Agile Progress Indicator Partnership
•• The nature of agile working environments •• Autonomous driving brings a new level requires new steering models of development complexity that can no longer be managed by individual players •• Progress indicators, such as OKRs (Ob- alone jectives and Key Results), should be used more as a compass to ensure movement •• Both technology companies and car man- in the right direction rather than a numeri- ufacturers benefit from complementing cal control on detail level each others’ skill sets and sharing develop- ment efforts
33 Area 1 refers to the product structure. The dictable in nature, such as the development clear interfaces. In the case of autonomous share of value added to the vehicle between of autonomous vehicles, require different driving, there are many unknowns for what hardware versus E/E and software is shifting approaches to success evaluation. Where constitutes the optimum technical solution. in favor of the second. Consequently, car the technology is new, the outcome uncer- Additionally, no single player in the industry manufacturers undergo major transforma- tain and the timeline unpredictable, classical possesses all the skills necessary to develop tions to refocus their core competencies KPIs often do not provide sufficient benefit the perfect solution. Google and Apple hold and build up expertise in those areas. in measuring and steering the progress of great expertise and talent in software devel- development. Instead, agile steering models opment and artificial intelligence, but do not Area 2 describes the change that is should focus on continuously measuring quite match the production infrastructure necessary on an organizational and work holistic progress and direction as compared and century old automotive engineering structure level. Structures and processes to performance at specific milestone dates. prowess of leading car manufacturers yet. that have proven successful for the devel- After all, agile development practices Car manufacturers possess the necessary opment of products with low shares of E/E support the ambition to cope with uncer- automotive expertise and infrastructure, and software are no longer ideal for the tainty by providing a new level of adaptivity. but lag behind with software capabilities. development of autonomous vehicles. Com- Upfront top-down planning, including metic- For this reason, companies are joining forces panies are increasingly replacing classical ulous project timelines, run counter to the in development partnerships and increas- waterfall structures with agile approaches. philosophy of agile development. However, ingly via mergers and acquisitions. The goal is to move away from long devel- a smart alignment between major top-down opment cycles, inflexibility and hierarchical project milestones (e.g. on a quarterly basis) command-and-control style management and bottom-up progress indicators, which practices and replace them with shorter de- show operational work advancements (e.g. velopment cycles, adaptivity, flat hierarchies on a bi-weekly basis), provide good orienta- and team empowerment. tion regarding the overall project status and direction. New work structures and development processes require new steering models, Finally, Area 4 addresses new forms of work- which is the focus of Area 3. Key perfor- ing with suppliers and partners. Traditional- mance indicators (KPIs) are an effective ly, car manufacturers have outsourced large tool to evaluate an organization's success shares of development work to suppliers at reaching targets. This applies mostly to in classic contract relationships. This is a situations that are predictable and linear. viable option when confronted with known Environments that are complex and unpre- technologies that can be partitioned with
34 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
architectural philosophy often still allocates software oriented "automotive technology specific functions to dedicated controller companies" with shorter release cycles for Product Structure: units). Simply put, this remains the industry software applications. The underlying change From Hardware to Software Focus standard today: One device, one function. To in product structure goes away from the "one For decades, traditional car manufacturers the same degree, it still holds true that tradi- function, one device" philosophy, towards have perfected their craft to build great qual- tional car manufacturers are cost optimizers a "many functions, one software platform" ity cars. Over the last 30 years, automotive more than innovative game changers. The approach, as shown in Figure 13. The E/E as well as software applications have current business model is to make money computing power required to process the gained great traction, along with the rise of by selling cars. New entrants follow a more huge amounts of sensor data, primarily from enabling technologies including computer radical strategy. Technology companies have cameras, radars and lidars in autonomous chips, the internet, etc. Initially, speed of inno- identified the value of data for their business vehicles, makes those cars supercomputers vation was slow in this domain compared to and leverage business models that support on wheels, as stated earlier. Thus, it makes modern standards. This is why it made sense this notion, such as Waymo's robo-taxi pilot sense to take advantage of this technology that a single control unit had the sole pur- program in Phoenix, Arizona. potential and use it as a central source for pose of powering a single function (of course, data processing. this is a simplified explanation because Traditional car manufacturers have identified several E/E functions are commonly execut- their need for change and are drastically ed across several ECUs. However, the E/E investing in becoming more agile, more
Figure 13 – Shift from hardware to software focus Traditional Hardware-driven Philosophy New Software-driven Philosophy
One Function Many Functions » One Device » One SW-Platform
35 The aforementioned changes do not only approaches. The main difference lies in the continue to become more connected and apply to hardware, but also to the way soft- software development sequence: While autonomous. The most prominent example ware is applied (Figure 14). Up until recently, waterfall follows a sequential path from of a company that is already using RSUs there was no demand for regular software conception to deployment, agile has an successfully is Tesla, which regularly releases updates. Control units were flashed during iterative approach where potentially ship- updates to its fleet to improve the autopilot production and in most cases never saw an pable software increments are developed function, battery range, or others. Other update until the end of the car's product according to a minimum viable product prominent players, including traditional car lifecycle. Software was designed under the approach. This is relatable to the way smart- manufacturers such as BMW and Daimler, premise of avoiding errors at all costs, and phone apps are created today. The updates have gained substantial experience with therefore required substantial lead and regularly released to app stores are product RSUs, not least because of their car-sharing development times. increments resulting from a sprint (in Scrum, fleets. a time-box of 4 weeks or less during which In agile software development, software a potentially shippable product increment is quality is of paramount importance as well, developed). Remote software updates (RSU) oftentimes even more so than in waterfall in automotive gain in importance as vehicles
Figure 14 – Shift from hardware-driven to cloud-driven software updates
Hardware-driven Software Updates Cloud-driven Software Updates
App Store
Remote Software Update
Upfront planning, updates after years Updates according to sprint cycle
36 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
»Companies in every industry need to assume that a software revolution is coming.«26 Marc Andreessen (Cofounder and General Partner, Andreessen Horowitz)
37 Figure 15 – Traditional waterfall vs. agile software development
Work Structure Traditional software development: Separated structure Agile software development: Integrated structure follows Product Structure As mentioned, traditional software devel- opment follows individual development Software component Software Software Feature Team 1 Cross-functional feature stages sequentially. From an architectural team incl. component Component Component team with end-to-end standpoint, building blocks are divided into lead with local respon- 1 4 responsibility software components, which are taken sibility Feature Team 2 care of by component teams (compare Software Software Figure 15). Since development times can be Component Component extensive for certain software components, Feature Team ...
2 … Community 1 Community 2 Community … Community m this approach requires meticulous upfront
planning: First, to enable teams to work Overall System in parallel with clearly defined interfaces, Software Software Feature Team n tasks and responsibilities; second, to reduce Component Component dependencies to a minimum, as queues 3 n have an exponential impact on cycle times. In complex systems, dependencies cannot Component Feature Backlog- be avoided every time, which increases Plan-driven Teams Cross-feature Teams driven coordination efforts and complicates system Dependencies between alignment via integration. Owing to this structure, water- components cause Function- End-to-End Command- self-organized specific Servant Development fall development is characterized by local queues, which have an and-Control communities Leadership optimization of single software components exponential impact on Leadership due to silo development. Operational and cycle times Local organizational structures are detached and Global Optimi- require coordination teams. Management Optimization zation Integrated follows a command-and-control style Function- operational & specific leadership, because technical decisions are organizational Development Waterfall Agile made top-down. Structure Iterative Sequential (MVPI Approach) Team Individual Respon- Cross- Responsibility Silos sibility functional
I Minimum Viable Product Approach 38 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
Traditional software development: Separated structure Agile software development: Integrated structure
Software component Software Software Feature Team 1 Cross-functional feature team incl. component Component Component team with end-to-end lead with local respon- 1 4 responsibility sibility Feature Team 2 Software Software Component Component Feature Team ...
2 … Community 1 Community 2 Community … Community m Overall System Software Software Feature Team n Component Component 3 n
Component Feature Backlog- Plan-driven Teams Cross-feature Teams driven Dependencies between alignment via components cause Function- End-to-End Command- self-organized specific Servant Development queues, which have an and-Control communities Leadership exponential impact on Leadership cycle times Local Global Optimi- Optimization zation Integrated Function- operational & specific organizational Development Waterfall Agile Structure Iterative Sequential (MVPI Approach) Team Individual Respon- Cross- Responsibility Silos sibility functional
I Minimum Viable Product Approach 39 This is in stark contrast to the characteristics velopment are fundamentally different from embodied by agile software development, the ones in the automotive industry. The car which in its core embraces self-organization industry is highly sophisticated and consists on team level, pushes decision-making of an advanced network of manufacturer, down to the lowest possible hierarchy level supplier and partner relationships. OEMs and fosters servant leadership. Teams outsource large portions of development are cross-functional and assume end-to- work to suppliers, which creates depend- end responsibility for a product feature encies and the need to define and manage (i.e. minimum viable product increment). clear interfaces. In addition, we are talking In analogy to the most common agile about embedded software development framework, Scrum, teams are by definition with significant hardware shares. These feature teams; development is consequently circumstances pose new challenges to agile oriented toward the highest customer value working models, which have their origins and prioritized via a backlog of development in pure software development. In order to items. Consequently, the organization cope with such high levels of dependency continuously strives to achieve a global op- and hardware shares, companies have to be timum. A perfect feature team is able to do willing to continuously challenge the status all the work necessary to complete a feature quo and adapt where necessary. When (backlog item) end-to-end. Cross-feature considering introducing an agile working alignment is self-organized and all feature model, you should ensure that not only is it teams work on a common software re- compatible with your overarching product pository. Operational and organizational development process, but also meets the structures are integrated and streamlined demands posed by supplier relationships, to focus on value-adding activities while hardware development and computing eliminating overhead. power constraints. Bill Gates famously coined the phrase: "Intellectual property Many companies view agile as the holy has the shelf life of a banana."27 If you want grail for securing innovation leadership, to create the future, you have to innovate which can lead to massive transformation as fast as your competition, at the very programs, sometimes without taking the least. Becoming agile and adaptive can help necessary time to analyze and evaluate the achieve that goal, but you need to be smart full spectrum of implications. Agile (soft- about it. Agile transformations in complex ware) development brings many benefits, environments such as the automotive but it has to suit the purpose and environ- industry constitute a fine line between sig- ment. The dynamics and challenges in pure nificant performance improvements and the software development environments such complete inability to act. as banks, insurance companies or in app de-
40 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
»Today, companies have to radically revolutionize themselves every few years just to stay relevant. That's because technology and the Internet have transformed the business landscape forever. The fast-paced digital age has accelerated the need for companies to become agile.«28 Nolan Bushnell (Co-founder Atari, Inc.)
41 You also need to use the right scaling strat- enough and therefore cannot sustain the egy. Autonomous driving development divi- momentum, support and positive energy sions typically employ several hundred em- required to drive the undertaking towards ployees for the software content alone. It is success. Top management as well as em- a tremendous challenge for any organization ployees may soon lose trust in the change if to change everything from the ground up. A it does not yield positive results. Therefore, real agile transformation not only changes we suggest starting with a pilot, a small agile the way developers work with one another; nucleus, where a group of highly motivated it fundamentally changes how the organ- people come together to function as spark ization operates, from the organization for the change and drive its initiation phase. structure, via the operating model, product Word about the pilot's first successes will architecture, verification and validation pro- soon spread into other groups or depart- cedures, to the culture, to name but a few. ments, who will then be eager to "go agile". For example, hierarchical barriers are bro- Figure 16 shows an example of a structured ken down, technical experts become tech- agile transformation roadmap, including nical leaders empowered to make technical some of the most crucial steps to consider decisions without the alignment obligation in an agile transformation (Deloitte's with their superiors, silos are eliminated and structured enterprise agile transformation replaced with strong cross-functional team playbook). setups - in short, interaction mechanisms, processes, structures and skill requirements change and need to be re-trained. For tradi- tional car manufacturers, this is a particular challenge owing to long-established and perfected processes, legacy systems and complex dependencies across departments, cooperation partners and suppliers.
Our experience with hundreds of agile transformation programs across the globe has shown that large enterprises are suc- cessful when they follow a structured agile transformation playbook. Most often, a "big bang" implementation, where the entire organization is flipped at once, culminates in a "big failure", simply because the organiza- tion is not able to change everything quickly
42 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
»Agile is more a 'direction', than an 'end'. Transforming to Agile culture means the business knows the direction they want to go on.«29 Pearl Zhu (Author of "Digital Master" book series)
43 Figure 16 – Deloitte's structured enterprise agile transformation playbook
Lean change: An agile approach to change Structured ...
Build
Work streams
Pivot Pursue MVC Establish agile vision and Define agile change and Stand-up agile change and Establish enterprise success criteria ops team ops team wide agile COEIV Operating model, alignment and organization design Measure Learn Define base operating Stand-up organization and Integrate and extend Refine and extend model and road map team structure across the enterprise
Measurement Framework Assess architecture and Componentization and Refactor architecture and Mature architecture Architecture and DevOps Operating DevOps capability DevOps strategy stand-up DevOps and DevOps model and alignment
Organization Architecture Define core agile SDLCI Refine SDLC method Apply and refine agile SDLC method design and DevOps method for pilots Agile Team Process and Practices transfor- Stand-up tooling Stand-up tooling Define agile tooling stack Setup agile tooling mation enablement team enablement team program Define training and Execute training Program Training Coaching Train pilot groups coachinng program (executive, management, POII, scrum master, RTEIII, team)
Team Training and Coaching process and Coach pilot groups Transition portfolio, programs, and teams by waves practices
Source: Deloitte 2018
44 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
... enterprise agile transformation playbook
Scale Pilot Refine Adopt
Establish agile vision and Define agile change and Stand-up agile change and Establish enterprise success criteria ops team ops team wide agile COEIV
Define base operating Stand-up organization and Integrate and extend Refine and extend model and road map team structure across the enterprise
Assess architecture and Componentization and Refactor architecture and Mature architecture DevOps capability DevOps strategy stand-up DevOps and DevOps
Define core agile SDLCI Refine SDLC method Apply and refine agile SDLC method method for pilots
Stand-up tooling Stand-up tooling Define agile tooling stack Setup agile tooling enablement team enablement team
Define training and Execute training Program Train pilot groups coachinng program (executive, management, POII, scrum master, RTEIII, team)
Coach pilot groups Transition portfolio, programs, and teams by waves
I SDLC: System Development Life Cycle III RTE: Release Train Engineer II PO: Product Owner IV COE: Center of Excellence 45 Typically, you need to overcome a number Figure 17 – Barriers and solutions in the agile transformation of barriers during the agile transformation (see Figure 17). First, humans have the ten- i rri rs o utions dency to resist any kind of change in culture, structure and roles within a company. You reate a mutua understandin of Human resistance to can counteract this resistance by creating a the chan e add new competences chan e in cu ture structure and ro es mutual understanding of the change, adding and tr it out in a pi ot new competences and trying it out in a pilot. Next, there is a lack of open communication. Instead of forcing a form of communication reate transparenc throu h sharin onto employees, create transparency, e.g. Lac of open communication information and wor in in pairs through sharing information, avoiding information access restrictions or working in pairs. Especially large, established corpo- Fai fast earn fast rations are often too risk-averse. In order to ein too ris a erse Fai ure is success if we earn become agile, you need to adopt a fail fast, from it Ma co m For es learn fast mentality, because "failure is suc- cess if we learn from it"30 - Malcolm Forbes. Another crucial aspect is the often seen lack Pro ide a so id mandate to of leadership buy-in. If there are no senior Lac of eadership u in leaders backing the agile transformation, mana ers and stron eadership this could lead to failure. You have to provide a solid mandate to managers and strong leadership support if you want the transfor- An gi tr ns orm tion re uires a ot sis riv n ori nt focus and imp ementation mation to be sustainable. Source: Deloitte 2018 “Agile 101: Discover the agile ways of working” Naturally, agile working models do not respond to the same steering mechanisms as waterfall approaches. The following paragraph discusses the differences in more detail.
46 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
New Steering Model i rri rs o utions Classical KPI-driven reporting systems reate a mutua understandin of often follow an underlying traffic light logic. Human resistance to the chan e add new competences Holistic top-down defined project plans with chan e in cu ture structure and ro es and tr it out in a pi ot detailed milestones and deliverables for each single work package (or even on the task level) are still considered to be the holy grail. On the other hand, a few months after reate transparenc throu h sharin Lac of open communication implementing a steering model, the mile- information and wor in in pairs stone-related progress tracking with classic KPIs usually indicates a (negative) deviation from the originally projected timeline and Fai fast earn fast consequently leads to red traffic lights at ein too ris a erse Fai ure is success if we earn aggregated project tracking overviews. Of from it Ma co m For es course, the underlying reasons range from bad planning accuracy and over-optimistic assumptions to execution problems. The Pro ide a so id mandate to Lac of eadership u in real problem here comes into play because mana ers and stron eadership of natural human behavior: Every single red traffic light seems to indicate a need for action. As a consequence, countermeasures An gi tr ns orm tion re uires a ot sis riv n ori nt focus and imp ementation are often initiated to fight red traffic lights individually without keeping an eye on the big picture, see Figure 18.
47 Figure 18 – From KPIs to progress Indicator
Traditional KPI-driven steering model Steering in an agile working model
If KPI shows a red traffic Product Owner (PO) prioritizes the product backlog based on a light, specific counter- comprehensive picture enabled by a holistic set of progress measure has to be indicators and feedbacks defined and triggered
If KPI target value is 1. 2. achieved, no counter 3. measure is needed
Bottom- Top-down up Self- Steering organized Trans- Concealment parency
Counter- Fail, Learn, measure Correct
Cross- Silo functional Progress Mentality KPI Mindset Target Indicator Indicator for Values Prioritization
Split Shared Holistic Responsibili- Risk of Local Responsibili- Interpre- ties (Individual) Optimization ties (Team) tation
Source: Deloitte 2018
48 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
Agile working environments require (objectives and key results) which combine progress indicators more as a compass to top-down planning (approx. 30%) with ensure movement in the right direction bottom-up defined OKRs (approx. 70%). rather than a numerical control on detail It is more important to spend some time level. The old mantra of "the more, the defining the right OKRs rather than having merrier" in terms of numbers of KPIs does too many, and they should follow some not hold true anymore - and in fact it never simple rules: Define SMART goals (specific, really did. Likewise, agile working principles measurable, actionable, relevant and time- should not be used as an excuse to avoid ly). Furthermore, make sure that there is any type of top-down milestone planning. one responsible individual for each OKR or At the end of the day, project success in progress indicator. The tricky part is being agile environments is similar to sailing: If smart in aligning the big picture milestone you do not plan any course or direction plan with short- and midterm agile pro- before you set sail, you will not know where gress indicators and deriving reasonable you end up. Even Silicon Valley tech players holistic countermeasures in case of major from Intel to Google use so-called OKRs deviations.
»Short for Objectives and Key Results. It is a collaborative goal-setting protocol for companies, teams, and individuals. Now, OKRs are not a silver bullet. They cannot substitute for sound judgment, strong leadership, or a creative workplace culture. But if those fundamentals are in place, 31 OKRs can guide you to the mountaintop.« John Doerr (Venture Capitalist and Chairman of Kleiner Perkins)
49 2.1.4 Mastering New Technologies Cross-industry partnerships are an inevita- As mentioned above, autonomous driving ble prerequisite to mitigate these complex is one of the most complex development challenges and to close own technology challenges in the automotive industry. The blind spots. All major stakeholders engaged broad range of required skills and capabili- in the development of autonomous driving ties barely exists in-house at any traditional solutions have established or joined specif- OEM, supplier or tech player. The latter are ic cooperations or partnerships (Figure 19, well positioned when it comes to software exemplary selection only, this overview is development and agile working principles not intended to reflect the full extent of AD to achieve shorter development cycles and partnerships and investments). In addition time to market, but often lack the experi- to the lack of technological or process ence with industrialization and scaling a expertise, there are several other reasons real hardware business like building cars. to join forces. Reduced development costs On the other side, OEMs and traditional and risk sharing between partners are automotive suppliers often struggle with further important drivers for the emer- the transformation towards a new agile gence of those cooperations. Lastly, from product and software development system a topline perspective, a larger addressable with significantly shorter cycle times for E/E customer base and associated revenue and software-related functions. potentials have to be mentioned.
50 Autonomous Driving | Moonshot Project with Quantum Leap from Hardware to Software & AI Focus
Figure 19 – Development partnerships and strategic investments (exemplary selection, non-exhaustive)