7 Hard Things About Product Management for Autonomous Vehicles

Source: Identifying an Uber ATG Vehicle and what’s inside the Uber ATG self-driving system, Uber ATG

The emerging Autonomous Vehicles industry

The autonomous vehicles (AV) industry is an emerging field at the dynamic crossroads of machine learning, robotics, and automotive hardware development. AV companies have varying missions, but what they share in common is the desire to bring about a safer future that will save countless lives and increase human productivity by letting robotic drivers take the wheel. AV companies look forward to a future where humans cease to worry about moving themselves from point A to point B.

The role of Product Manager for AV Development

In an AV company, as in other domains, product managers own problem definition and solution prioritization. Given the aforementioned need to focus finite resources on the most impactful problems, they play a pivotal role in making self-driving cars a realty. Yet, outside of AV companies, it isn’t widely understood what PMs do in such a research-oriented context. I often get by aspiring AV PMs to talk to them about what product management for self-driving cars entails. This post aims to provide some insight into that question by surveying some of the key challenges product teams like mine face.

Self-driving car PMs must seek the truth

Because AV development requires applied research in robotics and AI, there is often high degree of ambiguity in both the exact problem and solution. This requires a truth-seeking mindset first and foremost. A product manager building self-driving cars must ask themselves:

  • What ought to be our objective?
  • How should we measure success?
  • What are the most impactful problems to solve now / this quarter / this year?
  • What are the relative merits of potential solutions?
  • What are the key risks and dependencies to achieving our roadmap?

Autonomous Vehicle PMs support a broad scope

There are some additional complications to consider. AV companies are not yet focused on large-scale commercialization. As such, their organizations are appropriately more research and engineering driven. This means that the nature of an AV PM’s job differs from their peers at Facebook or Google — or for that matter, Uber’s core ride-sharing business — in important ways. Product managers must not only guide the product development roadmap, but also enable and guide the R&D agenda.

Source: author’s analysis of LinkedIn data

Why Steve Jobs is not a good role model for AV PMs

In order to be successful, self-driving car product managers must master a deeply technical domain, balance intricate dependencies, and have a high emotional quotient (EQ) — especially humility — in order to lead world-class machine learning researchers and engineers. What’s more, given their sparse ranks, they cannot always provide the depth of guidance and breadth of coverage that PMs typically provide software engineering teams.

The power of EQ: humility, self-awareness and empathy

Instead, an AV product manager’s superpowers shine through in developing insight across various domains in order to innovate new autonomy capabilities. As a result, product managers must have bi-modal strengths: sufficient technical depth as well as as a deep well of humility and self-awareness.

Cultivating a safety mindset

Moreover, humility tends to better foster a safety mindset (imagine the implications of the opposite — let’s call it brazen self-confidence — on on safety). Product managers for AVs must help bring about advances in robotics while keeping in mind that their systems may not always work as designed or the designs themselves may be flawed. This awareness is necessary to anticipate and mitigate negative outcomes.

Challenge #1: Making robots that perceive the world

Source: PnPNet: End-to-End Perception and Prediction with Tracking in the Loop, Uber ATG

Working effectively with applied researchers

For a product manager, this means that an ability to work with researchers in an environment of rapid iteration and being able to prioritize the path to a production system is essential. This is one of the reasons I noted above that high emotional intelligence and humility are needed. Unlike, say, mobile or SaaS applications, the development process doesn’t always start with a PM defining a problem — it may be a researcher or production engineer coming up with their own set of experiments to be prioritized and tackled. Instead of following a single path, research teams are likely to run many parallel experiments, abandoning failures quickly and following up on successes with continually evolving approaches.

Building platforms to accelerate AI innovation

Product managers must also focus on the platforms needed to support these AI workflows. Just as designing a mold encapsulates the requirements of the resulting three-dimensional plastic object, AV platforms involve close collaboration to define the end product. For example, the capabilities of a machine learning platform will inform the rate at which experimental ideas and new datasets can be converted into better AI algorithms powering AVs. Similarly, sensor simulation is one of the AI frontiers that self-driving car researchers are pushing forward, in order to provide perception engineers with better ways to test their computer vision algorithms.

Challenge #2: Designing for real-time compute

Challenge #3: Processing massive volumes of data

AV companies generate immense amounts of data today

Current generation of self-driving cars can generate several terabytes of data per day. Depending on the specific configuration and sensors, the Automotive Edge Computing Consortium estimates a range of 0.4 TB per hour to 5 TB per hour. The data collected and stored includes data captured from multiple camera, radar and lidar sensors. Therefore a single car, driving 10 hours/day produces 4 TB per day at a minimum. (Separately, Intel also estimates 4 TB per day, so that seems like a reasonable and conservative figure to work with).

Product managers must be data and infrastructure savvy

One obvious challenge for product managers is to ensure the core platform infrastructure is robust and performant. Data must be frequently off-boarded from self-driving vehicles and processed to ensure ready access for myriad use cases. This includes processing the data to add metadata for analytics or log playback.

Challenge #4: Turning raw data into “Ground Truth”

Source: How do self-driving cars see?

Challenge #5: Defining metrics for a new industry

Source: Measuring Automated Vehicle Safety (RAND)
  • What is “good” driving?
  • How do we measure ML iteration productivity?
  • How do you set Pass/Fail thresholds for simulated tests?

Challenge #6: Serving as the “glue” across tech areas

Source: Typical Autonomous Vehicle System (ResearchGate)
  • In which cities/neighborhoods should our AVs drive?
  • How will changes to one part of the autonomy system impact another?
  • What novel technologies are coming down the pipeline from R&D ?
  • How can we design for safety as we bring together disparate systems/teams?

Challenge #7: Blending hardware and software

Source: NVIDIA

Final thoughts: Safety is always paramount

AV product managers face myriad challenges in order to achieve the industry’s shared purpose of a driverless future. A common thread you have noticed is that each one of these has an impact on safety, and PMs must keep safety in mind regardless of the particular problem at hand. There is no option to “move fast and break things” when transporting people in fragile containers at high speeds. Instead, AV organizations must “move fast with a safety-critical mentality.” Product managers can play a vital role in both the speed of innovation required and safety, by focusing the organization on, and developing solutions to the most important problems.

--

--

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store