FORMULA 110
Kris Jordan ↗
UNC–Chapel Hill / COMP110 / Fall 2026
In Formula110, students in their first programming course write the Python code that drives a simulated race car.
Follow the storyStudents write Python programs that drive autonomous race cars.
Formula110 is a physically based racecar simulation developed in-house by Professor Kris Jordan and Isabella (Izzi) Hinks for COMP110 at UNC–Chapel Hill. It was built for the course, with Carolina-blue cars and a track where students could test their code.
COMP110 teaches programming fundamentals alongside AI competencies. Students first encountered the Academy exercise after just three weeks of programming: variables, conditionals and arithmetic were enough to begin teaching a car to drive. The experience built toward live, head-to-head races between student controllers.
Sixty times a second, the simulator calls a student’s Python function with fresh measurements of wall distance, speed and the track ahead.
The function returns steering and throttle commands. The physics engine applies those commands to the car, then the function receives the next sensor readings.
Sensory inputs
wall_lidar.front_left_m and front_right_m measure meters to the barriers, 20° either side of the nose.odometry.speed_mps reports forward speed in meters per second.center_offset_m, heading_error_degrees and lookahead offsets.Driving outputs
steer
−1.0 / Full left 1.0 / Full right
throttle
0.0 / Coast 1.0 / Full gas
Braking uses the same throttle input. Negative values brake a forward-moving car, then drive it in reverse once its speed is low enough. The actual API accepts throttle from −1.0 to 1.0 and has no separate brake field.
students wrote code
to control a car
students qualified for Formula110
controller decisions each second
The homework develops one controller through four levels, each using another programming idea from class.
Students begin the individual Academy exercise by driving a lap themselves. They then write the decisions that will let the car complete one on its own.
Each level stays in its own file. Students race the next version against the previous one, changing one thing at a time and checking whether it helped.
Read two forward wall sensors. Use conditionals to steer away from a nearby barrier and complete a lap at low throttle.
Compare the speed sensor with a chosen maximum. Request more throttle below the limit and ease off at or above it.
Replace the on-or-off speed choice with an arithmetic expression. Throttle changes continuously as speed changes.
Read the documentation and design a strategy using processed camera inputs. Replace the two forward wall sensors; speed readings remain available.
if sensors.wall_lidar.front_left_m < 4.0:
steer = 1.0
else:
if sensors.wall_lidar.front_right_m < 4.0:
steer = -1.0
else:
steer = 0.0
throttle = 0.16
return RobotCommand(
throttle=throttle, steer=steer
)
This Level 0 example steers right when the left wall is close, left when the right wall is close, and straight ahead otherwise. A low throttle gives it time to respond.
Formula110 brought the students’ controllers together for an evening of racing in Genome Sciences 200.
COMP110 instructors Kris Jordan and Isabella (Izzi) Hinks hosted the event, with races drawn from the qualifying categories. The programs students had developed and tested now drove alongside one another on the big screen, while classmates followed the field and cheered.
Several races introduced circuits the controllers had not encountered in qualifying. The evening’s finale sent ten cars around the Bahrain circuit for three laps, testing how their driving strategies handled a different course.
COMP110 students watch their programs race in front of classmates.
Robotics and self-driving cars
Provost Magnus Egerstedt, a roboticist, joined the event to talk about robotics, AI and self-driving cars. He connected the students’ controllers with the problem of using information about a machine’s surroundings to decide its next action.
From a driving decision to autonomous systems
Egerstedt’s work includes autonomous vehicles, environmental monitoring and groups of robots moving together. These examples provide research context for the control problems students encountered in Formula110.
With Georgia Tech’s Sting Racing team, Egerstedt helped develop the control architecture for Sting 1, a Porsche Cayenne entered in the 2007 DARPA Urban Challenge. The project brought sensing, planning and control together in a full-size vehicle.
Read the team’s control-architecture paper ↗ Photo: Dave Wooden / Georgia Tech · Source · CC BY-SA 4.0. Resized and converted to WebP; full frame retained. Photo adaptations shared under the same license.
SlothBot explores how slow, energy-conscious movement can support long-term environmental monitoring. This early prototype travels along a cable; later versions carried sensors into the tree canopy.
Read about SlothBot’s field testing ↗ Photo: Gennaro Notomista / Georgia Tech · Source · CC BY-SA 4.0. Resized and converted to WebP; full frame retained. Photo adaptations shared under the same license.
Simulation lets students test a controller, watch how it behaves and revise it. They can repeat a run, compare two versions and investigate a mistake without damaging a physical car.
The research examples above show related control problems in physical robots. Formula110 gives students a place to practice the basic process: read measurements, choose an action and test the result.
COMP590 / Honors AI Engineering
Students in COMP590 Honors AI Engineering presented their own approaches to Formula110.
Their projects explored ways to train a driving strategy from experience and combine learned behavior with programmed decisions. The demonstrations included learning from driving data and using the situation ahead to decide when to push for speed or take less risk.
COMP110 and COMP590 students were using the same simulation at different stages of their studies. The presentations showed how more advanced methods could be applied to a familiar driving problem.
Caleb Han and Mason Mines built an interactive website for their COMP590H project, showing how different driving approaches behave on the track.
Interactive website by Caleb Han and Mason Mines / COMP590H
Explore the student project ↗Six races, with fields selected from the qualifying leaderboard.
These videos replay the event lineups in new simulator runs.
Videos start muted. Use the player controls to turn on sound.
01 / Low g-forces
Sips Tea selected controllers with the lowest accumulated horizontal g-force load over a completed lap. The score includes acceleration, braking and turning, favoring a smoother ride. Qualifying averaged each controller’s lowest-scoring completed lap from each of five starting positions.
Watch Sips Tea on YouTube ↗
02 / High g-forces
G’s Going Crazy selected controllers with the highest accumulated horizontal g-force load on laps with no wall contact. Acceleration, braking and turning all contributed. Qualifying used each controller’s highest-scoring eligible lap from each of five starting positions.
Watch G’s Going Crazy on YouTube ↗
03 / Damage and survival
Hits Different selected controllers with the most total damage across five qualifying runs. Every run still had to last the full 30 seconds, make forward progress and finish with the car below 100% damage.
Watch Hits Different on YouTube ↗
04 / Fastest laps without braking
Gas Locked In selected the fastest completed laps with no applied braking. Coasting was allowed. Qualifying averaged each controller’s fastest eligible lap from each of five starting positions.
Watch Gas Locked In on YouTube ↗
05 / Fastest clean laps
Clock It selected the fastest completed laps with no increase in damage and no wall contact. Qualifying averaged the fastest clean lap from each of five starting positions. This field raced on the original qualifying circuit.
Watch Clock It on YouTube ↗
06 / Three laps on a different circuit
The finale selected the leading eligible controllers from All Spawns, No Crumbs, which ranked progress using each car’s weakest result across five starting positions, after recovery penalties. Controllers using concepts beyond COMP110 were excluded from this field. The finalists then raced three laps around Bahrain.
Watch the Bahrain Grand Prix on YouTube ↗
Event photography: Jeyhoun Allebaugh / University Communications and Marketing.
Formula110 began in the summer of 2026 and continued to develop as students used it that fall.
The simulator brought vehicle physics, sensors and race rules together around a small Python interface. Cameras, timing displays and the opening sequence were added as the project grew into a race-night event.
01 / June 27–28
The early prototypes established a car and a road to drive on. Lighting, barriers and painted kerbs added detail to the course, while overhead and following cameras offered different ways to see the car’s behavior.
02 / Summer 2026
Early tests put more than one car on the circuit. The car’s bodywork and paint also developed through the summer, taking on the open-wheel shape familiar from Formula racing.
03 / July to September
By July, Formula110 had a name and a Carolina-blue-and-white argyle strip painted across the asphalt at the starting line. September refinements carried that identity into the car, logo and opening sequence.
04 / Early September
Academy organized the driving problem into a sequence of programming tasks. Each level introduced another way to use the available sensor readings, with a previous version of the controller available for comparison.
Sensor readings
↓
Python controller
↓
Driving command
↓
Next sensor readings
05 / Late September
A close camera view helps explain what one car is doing. A full race also needs a way to follow the rest of the field. Split-screen cameras and the timing display let viewers track individual cars without losing the race order.
06 / Preparing for Race Day
By the final preparations, the simulator could bring a field of student controllers onto the grid and follow them through a race. Bahrain provided a different circuit for the finale, extending the challenge beyond the track used in qualifying.
“I engineered and co-created Formula110 with today’s frontier AI tools, specifically OpenAI Codex. It would not have been possible without them.”
For me, Formula110 is an optimistic example of AI helping us build more engaging learning experiences on faster timelines. An idea for a programming exercise became a simulation students could drive in, experiment with and race together.
— Kris Jordan