“`html

In the weeks preceding this year’s IndyCar Java House Grand Prix of Monterey, eight racing automobiles raced around one of the most renowned road tracks in the United States: the WeatherTech Raceway Laguna Seca. Yet, there were no famous drivers in the seats. These were self-operating, autonomous cars crafted by university teams and supported by AI, maneuvering around bends and speeding down Laguna Seca’s straight sections, occasionally reaching velocities nearing 150 mph. The newest squad in the group, the Caltech Racer, delivered what analysts described as a remarkably swift performance.

The event on July 24 was a segment of the Indy Autonomous Challenge (IAC), an ongoing multiyear initiative launched in 2019 by the Indianapolis-based nonprofit Energy Systems Network “to motivate university students globally to conceive, innovate, and demonstrate a new generation of automated vehicle software to operate fully autonomous racecars and inspire the forthcoming generation of STEM talent,” according to the IAC website.

Currently, there are 10 active squads in the challenge, representing universities and public-private collaborations worldwide. The IAC provides each team with a “drive-by-wire” IAC AV-24 vehicle equipped with a variety of sensors. The major distinction from team to team resides within their computing architectures—the sophisticated control and autonomy algorithms and intelligence systems operated by computers that enable the race cars to observe their surroundings and make decisions regarding acceleration, braking, and navigation as they race around the track.

Caltech joined the challenge in July 2024, acquired its vehicle in September, and made its racing debut at an IAC event at CES (Consumer Electronics Show) in Las Vegas in January 2025. The team comprises more than a dozen students, staff, and faculty members, alongside research scientists from NASA’s JPL, which is overseen by Caltech, and has greatly benefited from the technical input of one of its funding partners, Beyond Limits.

For Caltech, the endeavor began as one of the latest projects of the Institute’s Center for Autonomous Systems and Technologies (CAST). Engineer Soon-Jo Chung spearheads the effort with former CAST director Mory Gharib (PhD ’83), and Fred Hadaegh, research professor in aerospace at Caltech. CAST has a history of developing systems that advance the frontiers of autonomous control, having created robotic systems such as flying ambulances and multimodal robots, as well as control systems like one designed to keep damaged unmanned aerial vehicles in flight and another to assist spacecraft in self-correcting during emergencies.

Hadaegh, also a former chief technologist at JPL, remarks that it has been exhilarating to witness Caltech launch CAST and make strides toward crafting intelligent, robust systems. “For years, we envisioned machines that could adapt on their own. Now, that vision is a reality,” he states. “We’ve developed and tested the capability to identify, diagnose, and autonomously recover from failures in real-time—even amidst competition—on unmanned vehicles, without human involvement. This is not automation; it is authentic autonomy.”

Chung, Caltech’s Bren Professor of Control and Dynamical Systems and a senior research scientist at JPL, clarifies that the IAC serves as the ultimate testing environment for the types of algorithms his team develops. “The guiding principle of my group is truly to find a balance between theory and practical application,” he states. “When we formulate a new approach to autonomous planning or perception-based navigation, for instance, we want to ensure that those algorithms can be verified in real-world conditions. Autonomous racing provides a realistic controlled atmosphere where students and researchers can examine and subsequently refine their techniques. Once you’re dealing with that kind of speed, it amplifies any challenges faced—that’s exceedingly unpredictable. It’s a potent motivation for us to consider the discrepancies between our laboratory work and the requirements in real-world scenarios.”

Gharib, who holds the Institute’s Hans W. Liepmann Professorship in Aeronautics and Medical Engineering, adds that the IAC effort presents an exceptionally demanding challenge. “IAC functions as a proving ground for Caltech graduate students to showcase their technical prowess, problem-solving skills, and relentless determination,” he remarks.

The Scene at Laguna Seca Raceway

The July contest marked the inaugural IAC event on a US road course instead of a standard oval track. (Two prior events have occurred on the Monza F1 Circuit in Italy, which is also a road course.) The WeatherTech Raceway Laguna Seca is recognized as one of the toughest courses globally, featuring a notorious left turn known as “The Corkscrew,” which includes a five-and-a-half-story drop over less than 500 feet of track. Besides such descents and sharp turns, the cars in the contest also faced pedestrian bridges in various locations across the course—elements that compelled the autonomous vehicles to employ innovative prediction models during moments when they briefly lost GPS signals that typically assist with navigation.

During the time trials, a single team would signal its vehicle to enter the track by pressing a button. From that moment onward, each car was essentially left to itself to complete the race. Following one warm-up, the “out lap,” the vehicles could spend up to 12 minutes attempting to outperform their rivals by achieving the quickest single lap around the 2.238-mile circuit. In the meantime, the operational team could only transmit simple commands to modify settings during the final straightaway at the conclusion of each lap—a window that lasted at most 10 seconds.

However, Thomas Berrueta, a postdoctoral researcher in Chung’s Autonomous Robotics and Control Lab and the team lead for Caltech Racer, clarifies that by the time the team arrived at the paddock on race morning around 5:30 a.m., everything had “more or less” been established. “Our code is set. We’ve tested it in thousands of simulations. We’ve trialed it on the hardware at the track. We’ve determined a racing strategy,” he explains. “There’s an overall silence in the atmosphere that lingers as you’re awaiting the moment you’re summoned to the control center to actually conduct your race.”

On race morning, the orange, white, and light blue Caltech Racer (featuring two plush toys riding on top: a beaver and a Pokémon Psyduck) was the sixth to execute its time trial based on the team’s impressive showing in qualifying runs. Berrueta and two other team members were situated in the control center, fixed on spreadsheets that outlined the car’s preprogrammed race protocol and scrutinizing live telemetry data from the vehicle. Other teammates monitored the vehicle from the pit lane, ensuring it was functioning as anticipated. Berrueta pressed what he referred to as a launch button, and Caltech Racer took off, adhering to what was intended to be an assertive racing strategy.

“The racing strategy on my laptop essentially consists of a sequence of racing lines or specifications for behavior that the car should aspire to execute.

“““html

to adhere to on a lap-by-lap basis,” Berrueta states. “The vehicle strays from these criteria, moment by moment, consistently, but they steer its general performance.”

“A significant factor of this race is that when you elevate the velocity on the turns, that will push the vehicle to the limits of its grip and friction thresholds, and if you surpass that, it will skid or lose control,” Chung clarifies. “Thus, it’s a balance: You want to optimize the cornering speed while ensuring the vehicle remains secure and does not skid excessively. We truly expanded the limits regarding what our autonomous technology could safely accomplish.”

Indeed, by being assertive, the Caltech Racer showcased the quickest first and second lap durations observed in the contest up until then. Regrettably, the vehicle spun out on the third lap in the notorious “Corkscrew” and collided with a wall. This resulted in a red flag that, according to race regulations, required the forfeiture of Caltech’s two fastest laps.

Ultimately, PoliMOVE MSU, a coalition from Michigan State University and Italy’s Polytechnic of Milan, seized first place, Purdue University secured second, and the Korea Advanced Institute of Science and Technology obtained third.

Although disheartened by the final result, the Caltech team found motivation in its initial laps and is utilizing the race data to improve its strategies.

“The Caltech team has made significant advancements—developing solutions to problems that have taken other universities years to solve—and continues to advance the boundaries of autonomous vehicle innovation,” Gharib states.

Race officials and commentators also acknowledged the team’s advancements.

“Observe Caltech,” noted announcer Shea Adam during the event’s broadcast. “They had a vehicle that was exceptionally competitive. They are a team I would anticipate being right there, vying for the win in the next competition we have.”

“I believe we witnessed some impressive performances today from several of our teams,” remarked Paul Mitchell, president and CEO of IAC and Energy Systems Network, during the awards ceremony following the race. “This contest features 10 teams that are exceedingly dedicated. And at any moment, these teams can ascend in rankings. … There truly were some remarkable laps executed by the team from Caltech—our latest participant. So, exceptional progress is being made.”

A Vital Stack at the “Wheel”

Heading into the contest, the Caltech vehicle had traveled over 1,200 miles in 2025—hundreds more than the team that recorded the second-highest mileage. But beyond “practice leads to improvement,” what distinguishes one autonomous race car from another? After all, each IAC vehicle shares the same four Bridgestone Racing Slick tires, the same modified IndyNXT chassis and engine, along with the same CPU (Central Processing Unit) and GPU (Graphics Processing Unit) processing data from an identical array of sensors. These include six color cameras, two radar devices, GPS sensors, and four LiDAR (Light Detection and Ranging) sensors that utilize lasers, rather than radio waves like radar, to construct 3-dimensional representations of a vehicle’s environment.

Within the cockpit, where a driver of an IndyCar would typically situate, each autonomous vehicle instead contains a large computer. That computer operates the team’s software stack—a blend of its motion planner, which can be regarded as “the intellect” of the vehicle, and the tracking controller, the software that issues commands to the car’s actuators that facilitate acceleration, deceleration, turning, etc.

To comprehend what the stack aims to achieve, consider one of the abilities that the finest human race car drivers have honed: The capacity to perceive how their car’s tires are engaging with the road at any particular moment. Drivers understand when to push their car harder versus when to ease off based on their intuition, or “feel,” for the car’s performance. Endowing an autonomous vehicle with something akin to that intuition is exceedingly complex and necessitates a sophisticated model, essentially a collection of equations, that delineate the physical constraints of the tires. This model is continuously refreshed based on the experiences of the robotic vehicle.

Berrueta elucidates that the Caltech Racer’s motion planner is perpetually formulating estimates of how far the car’s components can be utilized. For instance, he states, as the vehicle navigates a turn, it will compute and then recompute how swiftly it can move without losing control. “It’s developing a subtle comprehension of just how much the tires can handle under varying conditions. Concurrently, it’s creating a strategy so that the tracking controller can bring those constraints into reality.”

“The faster we go, the more intricate the physics become, and the more challenging it is to model precisely,” remarks Nikhil Ranganath (MS ’20), a graduate student in Chung’s lab and a contributor to the Caltech Racer team. “The fundamental challenge is that the vehicle, in order to achieve maximum speed, is constantly pushing against our uncertainty regarding the model’s parameters.”

That uncertainty is perpetually present because conditions are continuously shifting. For example, he mentions, the accumulation of wear on the tires during a single race will influence the vehicle’s traction on the road. Even elements such as dust gathering on the tires or the temperature and humidity of the racetrack will modify the vehicle’s performance.

“There are countless minor factors that can occur which impact your understanding of what your tires are capable of, making it very difficult to maintain a truly accurate model,” Berrueta states.

Being Bold, With Safety in Focus

And the ramifications of an inaccurate model can be severe for these million-dollar machines. If the vehicle loses grip or a wheel locks, as it did in “The Corkscrew” at Laguna Seca, it can spin out and crash.

“One of the fundamental obstacles of this project is that you are genuinely striving to accelerate, but ultimately, you’re operating with a real vehicle in proximity to real people, and you must ensure it remains safe,” asserts Ranganath. “One lesson learned in the process of conducting field robotics is how to embrace aggressive development while simultaneously constructing a system that is inherently safe.”

That is why the team asserts that autonomous racing serves as an enthralling testing ground for robotic algorithms. Not only must the vehicle’s code operate smoothly and at breakneck speeds, but performance and safety are simultaneously mandatory. If the team sacrifices a fraction of performance for added safety, the vehicle could easily fall behind in the race.

And since autonomous systems are gaining traction across society, the advancement of these algorithms is directly pertinent to a plethora of other applications—from the management of oil rigs and self-driving vehicles to the operation of remote spacecraft. “Autonomy is ubiquitous in our society,” Berrueta states. “There are virtually no limits when it comes to rapid responses to failures and shifting conditions.”

A Strength in Swift Online Adaptation

The Caltech Racer team has gathered extensive knowledge in the year since it entered the IAC competition last autumn. Although none of the team members anticipated being involved in autonomous racing, they are exhilarated to engage in the challenge.

“It is extremely uncommon in research to engage in something that is so directly applicable, with such tangible, real-world implications,” Berrueta states. “When I make a mistake in a single line of code, my vehicle may crash. The way in which your work has a direct physical effect that is irreversible in some respects is rare for a research endeavor. Typically, we don’t work with platforms that are this demanding, costly, and specialized.”

Looking ahead to future IAC events, the team indicates they are improving the integration of machine learning into the race car’s motion planning, control, and perception, enhancing its adaptability and responsiveness to real-time track conditions, such as tire temperature or dust on the track.

All in all, Hadaegh states the opportunity to bring genuine autonomy to race cars is extraordinary. “These platforms require speed, accuracy, and resilience in the face of uncertainty—and now we can deliver. The IAC has provided us with the ideal arena to showcase our autonomy and demonstrate that onboard technologies are not merely concepts, but breakthroughs with genuine real-world impact. This is where the future of autonomy is being crafted—boldly, and in real time.”

The Caltech Racer team is sponsored by Beyond Limits, a Caltech/JPL spinoff, through a collaboration with Aramco.

“`