Smart from start to finish: GM’s path to new processes and products
General Motors uses intelligent assembly lines to manufacture the intelligent cars of tomorrow
What if we could spend more time with our passengers instead of watching the road? What if we didn’t have to find parking spots but could have our cars drop us off and pick us back up after our shopping trips? What if we fell asleep in our seats and woke up the next morning 300 miles away at our destination? Self-driving cars have obvious advantages for us consumers yet also offer significant benefits for businesses: How much could Uber save if they were able to cut the cost for drivers? How much more productive could agriculture become if machines operated on their own, day and night? How much space could communities save if they did not have to provide parking close to people’s’ workplaces?
With decreasing cost of computing, staggering amounts of data being collected every day, and revolutionary advances in computer vision thanks to convolutional neural networks, the dream of autonomous cars have become closer than ever. Aside Silicon Valley darlings like Tesla and Google, virtually every car manufacturer and major technology company has a horse in the race in order not to fall behind. General Motors is considered one of the front-runners after its acquisition of autonomous car startup cruise and $500M investment in Lyft back in 2016. With an aggressive hiring goal of 700 engineers dedicated to the research and development of new technology, a cash injection of $2.25B from SoftBank, and sustained investment into cruise the company has made significant progress: Earlier this year, the company submitted their application to operate a fleet of self-driving vehicles under a ride-hailing business to the DMV. The service is expected to launch next year [1].
But for GM the appeal of machine learning lies not only within new product development: Whoever reaches the goal of the first fully autonomous car, will have to serialize its production and get ready for mass production (unless they decide to sell). GM has built cars for 110 years, selling just under ten million vehicles in 2017 alone [2]. And aside from decades of experience in the manufacturing space, the company sits on terabytes of machine data growing every day. With this staggering amount of repetition and data, they are in a pristine position to utilize machine learning for process optimization, decrease cost, and improve overall output.
First to introduce assembly robots in automotive manufacturing back in 1961, GM today operates 800 to 1200 assembly robots in a single plant alone [3]. To improve the utility of GMs assembly robots last year the company patented a technique in computer vision. Rather than hard-coding the coordinates and shapes of every part at any point in time in the assembly line, the new algorithm gives robots the eyes and brains to track and handle a plethora of parts in real time, fully on their own [4]. And to systematically improve overall quality and yield of all of their newer assembly systems, the GM research team submitted a patent earlier this year for a technique that combines machine sensor data during production with warranty repair data that gets collected once the part has left the factory in the postmarket. This technique lets GM identify ‘hidden’ indicators for quality across each and every part and identify those parts in need of rework before they ever leave the factory floor [5].
As the entire industry is sprinting towards the goal of fully autonomous driving, regulators are lagging behind in creating guidelines for safety and proper conduct, leaving companies with plenty of grey area. While McKinsey eventually estimates a reduction of traffic casualties by 90 percent through autonomous vehicles, the industry is starting in the red: the fatal crash between a self-driving vehicle operated by Uber and a cyclist in Arizona in June of this year has the public weary and politicians ready to condemn the nascent technology [6]. As GM prepares for the launch of their autonomous taxi-service in 2019 they will face increased pressure from investors to go to market. It will be challenging to maintain the highest accuracy of the system across a universe of real-world scenarios. So far the company has relied entirely on its own fleet to generate training data. However, as timelines are shortening and rare traffic situations are, by definition, rare, GM could consider a model similar to Tesla: Over-equip top of the line consumer vehicles with sensors usually reserved for fully autonomous driving and let the community capture sensor data [7]. At this point in the race it is not about conserving cash, but covering all your bases to provide the safest product possible.
Questions for classmates
In machine learning ‘human level accuracy’ is often the gold standard to beat. Given humans’ poor performance when operating vehicles and the often fatal consequences of the smallest mistakes, what is the appropriate standard for self-driving cars? How should the system be judged in scenarios that have no desirable outcome (e.g. someone has to get hurt – either passenger or bystander)?
Sources
[1] CB Insights, “46 Corporations Working On Autonomous Vehicles,” September 4, 2018, accessed November 2018.
[2] General Motors, 2017 Annual Report (Detroit: General Motors, 2018), p. 2.
[3] James Amed, “GM Prefers Smart Manufacturing to Industry 4.0,” Ward’sAutoWorld, June 22, 2018, accessed November 2018 via Factiva.
[4] Payton et al. Robotic device including machine vision. US 20160368148 A1, United States Patent and Trademark Office, December 19, 2017.
[5] Wagner et al. Automated stochastic method for feature discovery and use of the same in a repeatable process, US 20170205815 A1, United States Patent and Trademark Office, July 31, 2018.
[6] Abe Kwok, “Uber’s self-driving car death was caused by a distracted driver” USA Today, June 25, 2018, accessed November 2018.
[7] Bernard Marr, “The Amazing Ways Tesla Is Using Artificial Intelligence And Big Data” Forbes, January 8, 2018, access November 2018.
As you mention in the article, the lack of regulation is the main barrier for autonomous cars to deploy their potential. While statistical data generated through these years should be enough to prove that autonomous cars are significantly safer than human-driven ones in terms of the frequency and the severity of the incidents caused, I believe the primary challenge still relies on defining who is legally responsible for the consequences of an accident. Should car manufacturers, government and insurance companies share accountability or should car owners be responsible for possessing the vehicles and assuming the risk? I think that for the technology to be quickly deployed, big manufacturers, government and insurance companies should take the responsibility and bear the consequences. Otherwise, switching costs would be high and owners would require a reliable and robust track-record before deciding to shift to autonomous vehicles. Regarding the moral challenges, the MIT Moral Machine initiative is one example of how to replicate human behavior in really ambiguous situations. I believe that reproducing our more common reactions to no-way-out conditions is the best solution to overcome this challenge.
Both of your questions hit on the two most significant questions about autonomous vehicles, in my view. The second question is one that I find fascinating because it brings up the question of programming ethics into autonomous vehicles. Do you specifically pick to kill an elderly passenger over a child if one of the passengers has to be killed? In these instances, the computer scientists and engineers are essentially playing God in making such design choices. The interesting thing is that human drivers make these choices all the time; however they are often forgiven for “bad” split-second decisions, such as turning into another car rather than swerving to the side of the road. Fair or not, that same forgiveness most probably won’t translate to the engineering team behind autonomous vehicles because they don’t have to make a split-second decision but rather are spending years making such choices. Thus, in these instances, I can foresee car manufacturers having to take blame and repetitively explain the rationale behind these programming decisions as these unfortunate events occur in the early days of autonomous vehicles.