ROAR Academy at a Glance — Berkeley’s AI/Autonomous Driving STEM Program

6 min readAug 26, 2021
  1. Introduction to Berkeley ROAR

Advancing the fundamental theories and applications in AI and Robotics has been one of the central themes in the emerging AI market. The industrial consensus seems to point to a market breakthrough in offering good enough solutions in the next five years to replace human drivers via autonomous driving systems.

The Robot Open Autonomous Racing (ROAR) program was founded by Allen Yang and Shankar Sastry at the University of California, Berkeley, in 2019 to tackle several remaining pain points in our community’s effort to bring such AI systems to be fully autonomous and be more safe than human drivers. These pain points are summarized as the highlights of the ROAR program:

Extreme Lower Costs: In order to develop and offer commercial AI solutions, the costs of developing new algorithms such as the approach of deep learning has grown exponentially. Computer costs in autonomous driving and maintaining a large fleet of vehicles is simply out of reach financially for both small to midsize companies and leading research universities. Could you imagine the cost of crashing an autonomous car during testing?

ROAR addresses these issues by developing and open sourcing low-cost AI, autonomous driving software, and hardware reference designs at a fraction of the cost. In addition to making our software free and publicly downloadable on GitHub, our hardware designs cost less than $500 to create an autonomous driving-enable RC scaled car that can exceed speeds of 70 mph.

Extreme Performance: Our ROAR platform allows researchers and students to easily test extreme driving corner cases and collect large amounts of human training data due to its low cost and safety due to the use of 1/10 scale cars.

We further introduced AR/VR capabilities built right into the hardware and software stack to allow human drivers to remotely control the vehicle as if they were inside a real, full-sized vehicle. This beyond-line-of-sight visualization allows researchers to develop 1/10 scale test tracks to solve dangerous driving corner cases in a low cost and safe environment. This is a stark contrast to performing risky real-world corner cases on expensive vehicles in urban city streets.

Extreme Compatibility: For educational programs that teach AI, robotics, and autonomous driving applications, a very fragmented developer ecosystem has been a significant challenge for researchers and students alike. Specifically for learning how to fluently develop efficient algorithms in Python, ROS, Tensorflow, OpenAI gym, and full-scale autonomous driving simulators such as CARLA and then finally deploying algorithms on a vehicle compute and ECU/ESC units to directly control the vehicle’s throttle and steering, it is expected that each of the above tasks would take a college student at least one semester to master from zero experience.

ROAR solves this problem by introducing an autonomous driving virtualization layer that is developed in Python. The virtualization layer allows even beginner-level learners to focus exclusively on Python coding to effortlessly deploy their test algorithms to ROAR vehicle reference designs, other autonomous driving simulators, and reinforcement learning gym environments with zero code changes.

2. ROAR Competition

Since the dawn of the automobile industry in the early 1900s, motor racing has been one of the most exciting high tech games for both adults and teenagers alike who dream of owning and racing their own supercars. Many advances in the automobile industry were first thoroughly battle tested via motor racing games, such as active suspension, traction control, paddle shifters, and recycling of kinetic energy.

With the same mentality and concepts in mind, the ROAR program created their AI racing competitions based on the ROAR open-sourced software and hardware design. We want our participants to modify the reference designs with their own creativity and submit their work to be tested against others. The ROAR program keeps a public record of the results, and also encourages all participants to open source their software and algorithms to benefit the community. As such, all related technical reports are continuously published on the ROAR website.

Currently, ROAR sponsors three AI racing leagues three times a year:

V1 Series (Virtual Reality): This series pits contestants to demonstrate their racing skills by manually controlling their vehicles through the support of AR/VR beyond-line-of-sight technologies. In this mode, driver assist functions can be deployed to aid the performance of human driving.

S1 Series (Software Simulation): A Python-based racing simulation environment allows contestants to race their own developed autonomous AI agents. Contestants can fully train and test their AI algorithms without any vehicle hardware.

A1 Series (L3-L4 Autonomous Driving w/ Vehicle Hardware): The A1 Series is the pinnacle of the ROAR program, where user-developed AI agents will fully and autonomously control a ROAR vehicle to complete a challenging physical race track.

3. ROAR Academy and Student Survey in Summer 2021

Since its inception in 2019, ROAR has supported various undergraduate-level and graduate-level courses and research projects at Berkeley in the areas of AI, control, computer vision, and human computer interactions. The platform has accumulated a rich library of teaching materials for entry-level learners to start their journey in AI education with the use of our open-sourced hardware and software reference design.

In Summer 2021, ROAR expanded its K-12 outreach by offering an AI STEM summer camp for 9th to 12th grade high school students, called ROAR Academy. In total, three classes were offered during July and August to 60 students. The academy supported high school student’s exploration of engineering as a career path by providing an interactive overview of machine learning, reinforcement learning, and upcoming autonomous driving technologies. We are very pleased to report that based on the end-of-program survey, the new learning experience for our enrolled high school students was overwhelmingly positive and supportive. Students surveyed described the Academy experience:

Students described our program as: Interesting, Engaging, Challenging, Comprehensive, Fast-paced, Well-organized, Hands-on, Informative, Intensive

More than 90% of the students preferred the ROAR Academy to stay at the same length or be a longer program.

More than 83% of the students surveyed think the difficulty level was just right or slightly difficult.

More than 80% of the students surveyed after completing ROAR Academy would love to continue to participate or attend future ROAR-related projects.

4.Future Expansion: ROAR Ambassadors program

The Ambassadors program is a new project that aims to bring Berkeley ROAR AI Racing, Research, and the ROAR Academy experience to interested K-12 schools and colleges. We put our technical and financial support into the hands of expert ROAR high school students who have successfully competed in the ROAR AI Racing Competition. Essentially, students will create a ROAR student organization at their respective schools with the goals of advancing autonomous driving research in conjunction with ROAR.

In addition to providing technical and financial support, the ROAR program will sponsor activities at partnering schools by organizing AI/Autonomous Driving seminars and ROAR AI Racing competitions. Additionally, we will support each ROAR ambassador with a ROAR faculty member and one Berkeley graduate student.

For more details about all the projects, competitions, and information related to the Berkeley ROAR program, please visit




Dr. Allen Y. Yang is Executive Director of Berkeley FHL Vive Center. Previously he served as Chief Scientist of Fung Institute, CTO of Atheer Labs,