The Mind of Argo AI

Co-founder and CEO Bryan Salesky brings a realistic approach to the self-driving technology “mission.”

An Argo AI Level-4-capable Ford Fusion on test in Austin, Texas, late 2019. (Image: Argo AI)

Creating technology solutions that work well in one area but don’t scale out is a traditional problem in the robotics field. It’s a particular challenge for engineers developing self-driving vehicles, acknowledged Bryan Salesky, the CEO and co-founder of Argo AI. The Pittsburgh-based company is a leader in SAE Level 4 systems development, partnering with Ford and Volkswagen to put autonomous people-movers and delivery vehicles into revenue-generating road use in the next few years. Ford plans to deploy its first units in three U.S. cities in late 2021.

“It’s no small task,” Salesky asserted. “We’re on a mission to make roads safer at the end of the day, and to give people options as to how they get around” with greater convenience and affordability. But such inflection points “don’t just roll out overnight—it’s going to be gradual deployments in cities that can support the platform,” he said.

Bryan Salesky: “I tend to forget how much we’ve accomplished in three years.” (Image: Richard Kelly Photography)

Argo AI’s test fleet of L4-capable Ford Fusion Hybrids is collecting data in six U.S. cities that offer infinitely differing driving scenarios, weather conditions and road characteristics. “We’re building a self-driving capability that can function well in all those different environments,” he told SAE’s Autonomous Vehicle Engineering during a recent interview at Argo headquarters.

Public-roads testing is vital for AV development, Salesky asserted, due to the need for ample real-world test data. Argo AI engineers record the sensor information and re-use the data, replaying it in simulation, he explained. “This ensures the vehicle is learning new things from it, that the object-detection software is working and that nothing has regressed in capability.”

A software engineer, Salesky was part of the Carnegie Mellon University team that won the 2007 DARPA Grand Challenge, considered a watershed event for proving AV capability. As a Michigan native who remains closely involved with CMU’s robotics experts, he has a Midwestern vibe that Argo colleagues say leads him to approach the self-driving business methodically, one step at a time.

“Bryan’s a pragmatist,” observed Larry Burns, the former head of General Motors R&D and author of the 2019 book, Autonomy. “He knows the profound changes in mobility aren’t going to happen overnight.” Salesky and Pete Rander, Argo AI’s president, launched the company in 2016. “Because I’m constantly looking at the next challenge, I tend to forget how much we’ve accomplished in just three years,” Salesky said.

OEM alignments

Partnering with two of the largest OEMs in a fast-moving, tech-intensive development space would appear to be daunting for a young outfit to manage. Salesky chuckles at the suggestion. “I suppose there have been challenges,” he quipped. “But we chose to partner only with companies that are compatible with the mission we’re on and share our view about how these vehicles will be deployed. We have a common understanding with Ford and VW that there will be shared fleets of these vehicles that are maintained by a fleet owner.” Personal-use L4 vehicles may come eventually, he reckons.

Argo AI and its OEM partners are “aligned around the fact that we need durable hardware that is manufactured for automotive use cases,” he noted. “We’re aligned around the basic sensor and compute approach. And we’re also aligned around getting to an operation where a driver is no longer needed – but there are many, many steps and phases before we get there. As long as our partners are aligned around those basics, the rest of it goes relatively well.” He added that both OEMs “have been supportive in everything we could ask for.”

Argo AI engineers at work in the company’s Pittsburgh headquarters. Growing the pipeline of engineering and computer-science talent is a focus of the company. (Image: Argo AI)

Understanding manufacturing has proved to be an elusive asset for newcomers to the mobility sector, OEMs and tech suppliers alike. Argo AI has, according to Salesky, “an appreciation for manufacturing timelines and what it takes to build things with the quality, reliability and cost attributes that automotive consumers expect—and that the ‘science projects’ don’t have to deal with, and conveniently ignore.”

Argo AI develops and prototypes much of its own hardware. “We’ve had to make the choice to design a lot of stuff in house, then work with Tier 1s and other suppliers to get it made,” Salesky explained. “We innovate when a solution doesn’t exist in the market. We prototyped our initial compute stack here in house then worked with a supplier to have it made.” The traditional supply base is still catching up to all the driver-assistance demands, let alone Level 4, he added.

Princeton Lightwave, a lidar specialist acquired by Argo AI in 2017, is prototyping a lidar sensor that Salesky promises will deliver market-leading range—part of the sensor strategy in Argo’s 10-year roadmap.

Filling the engineer pipeline

Salesky expressed concern about sustaining the pipeline of engineers and computer scientists for the mobility industry. Long-range funding for the academic community — commitments of at least five years — is a priority that he is pushing within the industry. His eight years in CMU’s robotics program deepened his appreciation for the challenges that engineering schools deal with, in particular stable funding, obtaining well-curated data and accessing the latest test platforms.

“We want to keep academia motivated for solving some of these really hard problems in robotics and of course self-driving systems,” he said. Argo AI came up with two supporting solutions. One is 'Argoverse,' a data set comprised of slices of well-calibrated road data captured during road testing in various cities.

Argo AI offers this package to machine-learning researchers at universities, among them Georgia Tech professor (and Argo principal scientist) James Hays. And as part of the Argo AI-CMU research center for AVs, engineering and computer-science seniors have the opportunity to test-code on Argo vehicles on a test track.

“That benefits all of us in the long run,” Salesky said. “It helps to get students interested in the field and helps to motivate research in this area. Robotics is a huge field and we want to make sure that academics don’t get the impression that the self-driving challenge is solved.”

Technical and structural changes in the mobility industry offer Argo AI opportunities, including major OEMs’ transition to all-new electrical architectures based on centralized compute. The trend “is about integrating software in an intelligent way,” Salesky observes. “Not every piece of software from every single supplier needs its own ECU! It’s completely inefficient. And if you want to make even a slight modification it creates huge ED&T bills and schedule delays. This is where I cringe as a software engineer.

“The industry needs to learn to iterate faster,” he said. “Over time, we’ll see dedicated computing for specific tasks. I’m not smart enough to say how all this plays out with autonomous-vehicle technology; we make use of 100 percent of the computing capability of these systems. But speaking broadly, it’s going to become ‘deliver me the software that performs a function’ rather than the computer and overhead that currently comes with a relatively simple function.”

Salesky sees important changes coming to engineering teams in the next decade. Efficient development of software is already a priority. Software by its nature is a much more iterative component than hardware. He says ‘waterfall methodology’ doesn’t apply in that “you don’t follow the waterfall once and then declare success. It’s following that sequence (requirement-design-implement-test) very fast, over and over again iteratively, until it’s done.”

As technology evolves over the next decade, OEMs will “get smarter” about how to build software. Unique in-house processes will allow them “to iterate and to evolve in parallel with and separate from the very rigorous ‘drumbeat’ that the typical hardware stuff is on,” he said.

Engineers then will make small modifications so they can respond to customer feedback much faster, using over-the-air updates. Validation will be quicker—“and those small changes won’t come with a multimillion-dollar bill,” Salesky stated.

2020 is a super-critical year for Argo AI. While city-roads testing with its partners continues, company engineers are no longer building the building blocks.

“We’ve moved beyond that phase,” he reported. “Now we’re focused on understanding what the commercial deployment looks like. How to handle first responders in emergency vehicles. How to interact with customers. Understanding delivery use cases and learnings on the commercial-vehicle front.”

No small task, indeed. Further progress lies ahead for the young company and its partners.