Pocket Racer: An accessible autonomous racing educational platform

Abstract

Accessible autonomous racing can engage undergraduate students emersed in artificial intelligence (AI) and robotics to receive an education through the excitement of maneuvering sharp corners and overtaking opponents. In particular, the construction and programming of miniature mobile racing robots can facilitate head-to-head racing competitions for use indoors within classrooms and hallways. Unfortunately, existing racing platforms remain inaccessible for overtaking maneuvers in such confined settings because they are physically large or expensive due to the computational cost of enabling such high performance. In an attempt to address these issues, we present Pocket Racer, an open-source, pocket-sized racing robot capable of head-to-head racing within indoor environments in education, i.e. university hallways or classrooms. We demonstrate head-to-head autonomous racing with our Pocket Racer platforms, enabling high speed overtaking upwards of 15 km/h. Designed to be easily assembled with off the shelf components and a low-cost edge device (Raspberry Pi Zero 2 W), our Pocket Racer platform is made accessible through an open source website and dataset detailing build instructions. By making a pocket-sized head-to-head autonomous racing platform accessible for undergraduate students, our work hopes to further hands on education in the age of physical AI.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-49690-x.

Introduction

Accessible as instructional tools^1,5 in indoor settings, miniature mobile robots² are useful for artificial intelligence (AI) and robotics education. Unlike open-source datasets^3,4 which do not have physical embodiment, miniature mobile robots⁶ give undergraduate students a unique opportunity to physically interact with AI. Recently, miniature mobile robots have gained interest within scaled autonomous racing⁷ in course-based racing competitions. These robotic racing competitions have popularized within robotics education as high speed overtaking of opponents motivates undergraduate students for collaborative learning⁸. Robotic autonomous racing also promotes competition-based learning^9,10. As undergraduate students evaluate the racing robot’s performance, they can use lap time results¹¹ as a form of objective feedback. Racing competitions not only develop technical competence¹², but also foster soft skills like cooperative teamwork¹³ and professionalism¹⁴. Furthermore, studies show that even when students are not interested in robotics^15,16, they are motivated to learn when teaming up to challenge others i.e. head-to-head racing¹⁷. These racing competitions also invite teams to disseminate technical results¹³ and discuss related methods¹⁸. Finally, head-to-head autonomous racing competitions cut across diverse engineering disciplines¹⁸ and are jointly held with universities around the world^19,20, opening intercultural communication.

Another important aspect of an accessible racing robot is its low cost, as seen in existing educational platforms^21,22. These miniature mobile robots² are often below $300 USD²³ for instructional use, as price becomes a key deciding factor²⁴ for academic institutions with modest budgets²⁵. Finally, open-sourcing is also necessary as it allows widespread adoption without restrictive requirements for access. In particular, the platform must be replicable, easy to build with clear instructions² and use readily available, off-the-shelf components²⁶. As a result, from miniature mobile vehicles²⁶, swarms²⁷, to larger quadrupeds²⁸ and manipulators²⁹, many open source robots have effectively applied hands-on learning¹² to teach students about artificial intelligence and robotics within confined indoor settings.

Unfortunately, low-cost miniaturization of a high speed mobile robot is challenging as performance often entails expensive or large hardware. As an example, existing full scale [19] or 1:8 scale racing platforms such as AutoRally³⁰ require outdoor testing facilities with its use of a RTK Global Positioning System (GPS) despite speeds upwards of 96 km/h. Comparatively smaller, 1:10 scale platforms such as F1 Tenth¹¹ and MIT Racecar⁸ use 2D Light Detection and Ranging Sensors (LIDAR), requiring the need to have large external walls not found in classroom circuits. Among hobbyists, the 1:16 scale Donkey Car²² platform is closest in meeting such requirements, but still large in size due to the use of the Jetson Nano compute board or standard Raspberry Pi compute boards.

Other 1:18 platforms from the industry such as DJI Robomaster³¹, JetRacer³² or AWS DeepRacer³³ are seen in racing competition events. However, these are slow in speed (< 7 km/h). Furthermore, many industry-based platforms are closed source, or require proprietary tools for operation, limiting their accessibility for educational use. Thus, existing limitations of larger racing platforms suggest the smallest 1:28 and 1:43 scale platforms are better suited for classroom use. While a few 1:28 and 1:43 scale platforms exist, (i.e. Minicar²³, Chronos³⁴ ORCAcar³⁵), they currently rely on motion capture systems which are highly expensive for educational use. Finally, recent milestones in robotics suggest the use of miniature aerial racing drones³⁶, but also require 3-dimensional testing facilities that are large in size. Thus, there exists an unsolved opportunity in AI education to develop an open-source, low-cost, miniature mobile robot specifically designed for head-to-head racing within confined indoor environments.

On an algorithmic level, many existing systems for autonomous racing are not data driven. Instead, they are based on control³⁷ and optimization³⁵, which use a hierarchical system architecture (i.e. perception, navigation, localization)³⁸. These controllers are built on top of Simultaneous Localization and Mapping (SLAM)¹⁹ methods which utilize global maps of the environment. Other methods rely on high fidelity, nonlinear vehicle dynamics²³ models and tire models for trajectory planning to ensure safety during multi-agent overtaking maneuvers. While useful for full scale applications, these methods are unnecessary for head-to-head autonomous racing within AI education.

There has been increased efforts to develop data driven autonomous racing methods, which do not rely on SLAM or physics-based models. For example, a few data-driven methods exist for end-to-end autonomous driving which are based on Imitation Learning (IL)³⁹ or Deep Reinforcement Learning (DRL)³³. However, these vision based methods rely on a sim-2-real approach, often requiring the simulator to visually match the visuals of the final environment or requiring an inaccessibly large dataset for generalization⁴⁰. Such domain transfer requirements for learning makes sim-2-real approaches difficult to apply ubiquitously in hallways and classrooms for educational purposes.

Behavioral Cloning⁴¹ is a well known Imitation Learning (IL) based approach to end-to-end autonomy which directly corresponds images to actions with supervised learning. This algorithmic approach is simple to implement which makes it accessible to learn and open source for AI education. As a result, existing platforms such as Donkey Car²² utilize end-to-end autonomous driving via Convolutional Neural Networks (CNNs)⁴² and Long-term Recurrent Convolutional Networks (LRCNs)⁴³. Furthermore, recent strides in Large Language Models (LLMs) for visual tasks^44,45 with the Vision Transformer (ViT)⁴⁶ suggest the Transformer encoder’s inherent sequential ability can extend towards trajectory planning i.e. autonomous driving. Prior work⁴⁷ has shown that ViTs with large datasets generally have fewer inductive biases (higher bias towards features over textures). We benchmark these end-to-end driving algorithms in an attempt to identify their suitability for autonomous racing in the context of AI education.

Thus, we open source Pocket Racer, an accessible autonomous racing platform for AI education. Unlike existing scaled autonomous racing platforms, Pocket Racer is both low in cost and size but enables high performance, making it suitable for accessible indoor educational use. Our efficient end-to-end approach to autonomy leverages image as its only input, ensuring ease of use and widespread adoption within the educational community. As suggested in Fig. 1, Pocket Racer uses on-board autonomy on a low-cost edge device with only an on-board camera, allowing indoor use without external reliance on sensors (i.e. MoCap) or extrinsic landmarks (LIDAR walls). Such flexibility of use within indoor environments (i.e. hallways) makes Pocket Racer meet the different needs of teachers teaching students with project based learning (PBL). As a comparison study, we open source a suite of end-to-end models i.e. ViT⁴⁶, CNN⁴², LRCN⁴³ and open source a collection of racing datasets which can be used to verify algorithmic performance for real world scaled racing competitions. The highlight of our work is exemplified in the robot’s high speed performance, which reaches speeds upwards of 16km/h.

Fig. 1.

Pocket Racer: the accessible platform enables autonomous racing competitions.

Results

The Pocket Racer platform

The Pocket Racer platform aims to be accessible and reproducible through an open source website that lists build instructions and demo videos (Appendix B) of high speed autonomous overtaking. A dedicated Github repository also provides code for all benchmark models for Pocket Racer. The platform build took two hours for undergraduates to complete with no prior experience, while given access to the Bill of Materials (BOM). Thus, our platform allows undergraduate students to hold head-to-head autonomous racing competitions with its low space and cost requirements.

Being 1:28th scale, Pocket Racer is readily available for Artificial Intelligence (AI) education in university hallways and indoor classrooms. Due to its sub $200 USD cost and 3D printable components, it is accessible for most academic institutions. The competition ready robotic platform enables overtaking using a single monocular camera and a Raspberry Pi Zero 2 W compute board.

Algorithm benchmarks

End-to-End autonomous driving via behavioral cloning (BC), a form of supervised Imitation Learning (IL), is a popular method for autonomous racing because of its ease of deployment, making it suitable for AI education.Convolution-based deep learning methods, such as CNN and LRCNs, face limitations due to distributional shift or covariate shift for temporal tasks such as autonomous driving. However, for spatial tasks soley involving steering control, CNN and in particular, 3D CNN, show high levels of predictive capacity for the purposes of scaled autonomous racing. Fortunately, recent advancements with Vision Transformers, including SwinViT⁴⁸, show promise in overcoming limitations caused by covariate shift at the expense of requiring larger datasets during training.

Thus, in an effort to assess existing algorithms, we evaluate various model architectures such as LRCN, 3D CNN, and Vanilla ViT, using MSE loss, accuracy, Hellinger Distance (HD), and Earth Mover Distance (EMD) to quantify autonomous racing performance. The HD and EMD metrics were selected for accuracy analysis as it would show the distribution of steering across a sequential dataset, as demonstrated in Fig. 2. Table 1 shows the accuracy results of five different deep learning based algorithms obtained within the natural science division hallways of Seaver College, Pepperdine University.

Fig. 2.

Distributional loss benchmarks: the prediction and target values are binned for the pepperdine hallway testset to show the distributional loss of the steering values.

Table 1.

Model accuracies for pepperdine hallway (490k) vs. classroom oval circuit (101k) (best accuracy of 40 trials).

Architecture	Img stack		HD	EMD
CNN34	1	0.999/0.995	0.0124/0.0165	0.0022/0.0026
3D-CNN	2	0.999/0.998	0.0085/0.0065	0.0016/0.0019
LRCN35	2	0.95/0.994	0.037/0.025	0.014/0.0028
ViT40	1	0.85/0.994	0.11/0.026	0.023/0.0036
3D-ViT	2	0.93/0.994	0.063/0.019	0.016/0.0047

The accuracies were obtained by running each algorithm for 40 runs on the Pepperdine Hallway dataset and 10 runs on the indoor mat circuit dataset to obtain the best results. CNNs for both 2D and 3D based approaches achieved the best results on both datasets, with 99.9% R squared accuracy on the Pepperdine Hallway dataset and 99.8% R squared accuracy on the classroom circuit dataset. The 3D variant of CNNs showed slightly better performance results overall for all metrics, including distributional loss metrics such as the Hellinger Distance (HD) metric and Earth Mover’s Distance (EMD) metric.

Open source dataset evaluation

In addition to our real-world tests in the Pepperdine Hallways (490k) and Classroom Oval Circuit (101k), we collected datasets from 20 circuits total in the process of developing our Pocket Racer Platform. We open source this supplementary dataset which can serve to validate models trained by students before real world deployment. We evaluate 5 neural network architectures (2D-CNN, ViT, 3D-CNN, 3D-ViT, and LRCN) on our 20 open source driving datasets with 4 velocity profiles (Slow, Medium, Fast, Race). After training, each model was tested on unseen test laps within each dataset corresponding to 10

To characterize dataset difficulty, we define two complementary scalar metrics grounded in Lyapunov stability theory: control energy and path flow. Together they capture the minimax game between two opposing forces that govern autonomous racing: the controller that expends energy to stabilize the vehicle, and the environment that introduces disturbances driving the system away from equilibrium.

Lyapunov Energy Function. Let denote the steering and throttle control inputs at time t. We define the Lyapunov candidate function as the cumulative control energy:

1

where T is the total number of timesteps in a lap. A higher V indicates that the controller must exert greater effort to maintain stability, corresponding to more aggressive or difficult driving conditions.

Lyapunov Flow (Energy Dissipation Rate). The instantaneous flow is given by the discrete time derivative of the Lyapunov function along the system trajectory:

2

When , the system dissipates energy and trends toward equilibrium (stable driving). When , external disturbances inject energy into the system, destabilizing the trajectory.

Control Energy. The total control energy over a lap of T timesteps is defined as:

3

This measures the mean squared magnitude of the control inputs and quantifies the stabilizing force in the minimax game: the controller seeks to minimize energy expenditure while maintaining accurate trajectory tracking.

Path Flow. The path flow captures the rate of change in the control signal, analogous to control jerk:

4

This quantifies the destabilizing force: environmental disturbances (sharp turns, surface changes, speed variations) that force rapid corrections in the control signal, injecting energy into the system.

Minimax Formulation. The equilibrium between these two opposing forces is expressed as a minimax game over the Lyapunov function:

5

where represents the environmental disturbance at time t and is an attenuation parameter. The controller minimizes V by producing smooth, energy-efficient inputs (), while the environment maximizes V by inducing high path flow (). Stability in the Lyapunov sense is achieved when the controller dissipates energy faster than the environment injects it:

6

where is the decay rate. Datasets with high and high represent aggressive driving regimes where both forces are large, while datasets with low values of both metrics correspond to smooth, low-speed driving near equilibrium (Table 2).

Table 2.

Per-lap RMSE distribution (mean ± std, IQR, 95% CI) averaged across datasets.

Model	Mean	Std	IQR	95% CI
2D-CNN	0.2687	0.0483	0.0480	[0.2373, 0.3001]
ViT	0.2778	0.0467	0.0501	[0.2471, 0.3085]
3D-ViT	0.2940	0.0437	0.0487	[0.2651, 0.3228]
LRCN	0.2945	0.0477	0.0482	[0.2634, 0.3256]
3D-CNN	0.3435	0.0583	0.0524	[0.3052, 0.3818]

Model training and deployment

We measure the time required to converge the suit of deep learning algorithms on a GPU workstation (CPU: Ryzen Threadripper 1950x, GPU: x2 Nvidia RTX2080Ti x1 Nvidia RTX4070 Super). Our results showed that the training time of all benchmark models for both hallways (490 k dataset) and classrooms (101 k dataset) averaged within several hours, making it suitable for use in project curriculum as well as racing competitions (Fig. 3).

Fig. 3.

Open Source Dataset Characterization: (a) Control envelope showing 95% and 50% confidence ellipses of steering and throttle distributions across four driving styles, where wider ellipses indicate more aggressive control inputs. (b) Per-dataset control jerk magnitude on a log-log scale, with bubble size proportional to sample count and color indicating aggressiveness. (c) Control energy and path flow per dataset on a log scale, with median and interquartile range error bars.

Maximizing latency was critical in minimizing cornering overshoot during deployment. In an effort to analyze latency, we assess inference time across an average of 1000 samples, as presented in Table 3. Among the tested models, LRCN achieves the fastest speed at 97.3 Hz with 3D CNN being the slowest at 68.39 Hz. It should be noted that without a model running within the loop, the Sony IMX708 camera was able to stream images at 97.66 Hz. Finally, Table 4 shows the accuracy of the best performing 3D CNN model against different image sizes. The best of 10 runs were used, and 32 x 32 pixels were sufficient to achieve 0.999% R squared accuracy. For deployment, ViTs were selected with its ability to demonstrate robust predictive accuracies while maintaining high inference speeds.

Table 3.

Model inference speed (avg. of 1000 runs).

Algorithm	Hz
3D-CNN	68.39
LRCN	97.3
ViT	85.64
3D-ViT	84.34
No Model	97.66

Table 4.

Model accuracies for image sizes (3D-CNN) show that even a small size of 32 by 32 pixels is sufficient for autonomy

Image size		HD	EMD
32 x 32	0.999	0.0085	0.0016
28 x 28	0.986	0.0296	0.1009
24 x 24	0.916	0.0478	0.1405
20 x 20	0.835	0.0670	0.1614
16 x 16	0.861	0.1149	0.0219
12 x 12	0.948	0.0681	0.0142
8 x 8	0.963	0.0667	0.0119

Head-to-head overtaking

Pocket Racer was designed for autonomous, high-speed overtaking maneuvers to enable head-to-head racing competitions. Classroom experiments showed that the platform can reach speeds exceeding 10 km/h in small classroom environments. Within hallways, Pocket Racer demonstrated top speeds of 16 km/h for single-agent racing and 15 km/h during overtaking maneuvers. Undergraduate students independently collected 20 demonstrations of overtaking maneuvers for three different Pocket Racer platforms to enable successful autonomous overtaking. As a result, the lap times of the three student-deployed Pocket Racer platforms had mean values of 39.3, 45.3, and 33.1 seconds across 10, 6, and 4 consecutive laps in hallway environments, respectively, as shown in Fig. 4.

Fig. 4.

Single lap time comparison: the hallway circuit laptimes for three different pocket racer platforms were evaluated for their performance.

Discussion

This paper presents Pocket Racer, an open-source, head-to-head autonomous racing platform, which is designed for use in AI education. Pocket Racer shows readiness for use in AI education due to being open-source, low-cost and small in size. Finally, Pocket Racer is designed for head-to-head autonomous racing, outperforming other existing mobile racing robots.

We observed that the deployment speed of the algorithmic models needed to match the speed demonstrated during data collection in order to avoid corner overshoot or delayed turning. However, it remains difficult to directly attribute overtaking performance to a single quantitative metric. Overtaking success is strongly dependent on track geometry, particularly track width, which determines whether stable side-by-side motion is physically feasible. This challenge also reflects a broader limitation of imitation learning via deep learning, since model behavior is inherently constrained by the coverage and consistency of the demonstrated data. Power consumption tests showed that the compute board (Raspberry Pi Zero 2 W) operated at 1.1 W (0.21 A at 5.1 V) during idle and 1.6 W (0.31 A at 5.1 V) during model inference.

Pocket Racer’s high speed performance is demonstrated by the high prediction accuracies achieved across a suite of deep learning algorithms^42,43,46 specifically tailored for edge device deployment. While the use of end-to-end autonomous driving has been applied in various other autonomous racing platforms^22,32, they are costly¹¹ and/or large in scale³⁰, limiting its usage in classroom settings. As presented in Table 5, our work enables high speed autonomy at 1:28 scale with minimal sensor requirements, enabling its accessible use in classrooms and hallways unlike other existing high speed platforms. In particular, the multi-agent demo videos on our dedicated website demonstrate Pocket Racer’s readiness for educational use via competitions. The platform’s size-to-performance ratio, provides a unique advantage by allowing it to be integrated into undergraduate curriculum as part of capstone design projects. With the Pocket Racer platform acting as a Cyber-Physical System (CPS), students can develop a comprehensive education; they build, collect, train and race their own data-driven robots.

Table 5.

Comparison of existing platforms highlight the scaled speed of the Pocket Racer platforms are significantly higher than existing platforms.

Model	Size	Cost (USD)	Onboard autonomy	Speed (km/h)	Scaled Speed
AutoRally 49	1:5	9210		32	160
F1 Tenth 12	1:10	3628		27	270
MIT Racecar9	1:10	3383		12	120
BARC car50	1:10	500		14	140
HOUND 22	1:10	3000		25	250
MuSHR v3 5	1:10	930		24	240
Robomaster 51	1:12	459		12	144
Donkey car 23	1:16	200		11	176
JetRacer 52	1:18	360		12	216
DeepRacer 53	1:18	399		12	216
Minicar 24	1:24	77		1	24
Chronos 54	1:28	300		9	252
DuckieTown 2	1:30	150		1	30
ORCA car 55	1:43	470		12	526
Pocket Racer (Ours)	1:28	196		16	448

In addition to fostering engagement, the competition-based structure of Pocket Racer can be explicitly mapped to measurable educational outcomes in undergraduate robotics courses. Therefore, we propose the following course learning outcomes (CLOs) and associated measurable indicators:

1. Design Challenge. Students can propose innovative hardware and software design solutions to address performance limitations of autonomous racing platforms.

2. Implementation and Evaluation. Students can implement and test new designs in simulation and hardware, and evaluate them using quantitative metrics such as lap times, successful lap completion, and number of overtakes.

3. Effective Communication. Students can communicate technical progress and experimental results through group presentations and written reports.

4. Teamwork and Leadership. Students can collaborate effectively in teams, as measured through peer evaluation, milestone reviews, and project coordination.

5. Project-Based Learning. Students can integrate hardware, software, and experimental validation into a final competition that serves as a live demonstration of project outcomes.

Miniaturization of high speed autonomy was identified as a key challenge in Pocket Racer’s development. In particular, miniaturization required trade-offs⁵⁶ between latency, accuracy and ease of use. For example, hyper-parameter tuning of end-to-end deep learning models is challenging for students as it requires domain expertise⁴⁷ unlike other control based approaches³⁵, which are based in first principles. To offload such initial challenges faced by students, we quantify potential trade offs and provide code for a suite of algorithmic baselines on a Github repository (Appendix C). As a result, students are given the ability to accessibly use the Pocket Racer platform with room to refine baseline models during competitions. Our dedicated website allows new students to easily replicate the platforms with little to no prior knowledge.

In an effort to objectively compare the performance of Pocket Racer for autonomous racing against other existing mobile platforms, we scale the speed according to the platform scale³⁵. Results showed that out of all existing mobile robot platforms with autonomous capabilities only using onboard sensors, Pocket Racer displayed the highest speed in terms of autonomous performance. Pocket Racer is also smaller than most platforms, with the exception of the ORCA car³⁵ which does not offer on board autonomy, and DuckieTown², which do not have the vehicle dynamics for high performance autonomy. Here, vehicle dynamics refers to physical characteristics of the chassis that enable high-speed performance, i.e. drivetrain configuration (gear ratio), tire material and wheel dimensions. This performance also holds value for education in classroom environments as it can be achieved with a modest budget of 196 USD, which is similar in cost to a DuckieTown² platform. While higher in cost to the Cambridge Minicar²³, Pocket Racer does not rely on a external MoCap system allowing it for use even outside the classroom.

In addition to its high performance characteristics, Pocket Racer is evaluated for its readiness for head-to-head autonomous racing. While autonomous overtaking maneuvers are key in enabling autonomous racing competitions, existing platforms are not designed for multi-agent use within the classroom due to its cost¹¹ and comparative size³⁰. As a result, the few existing platforms designed for head-to-head racing competitions^8,11,22 remain constrained for use in large hallways or auditoriums and do not offer the flexibility of usage in classroom environments. Furthermore, as illustrated in Fig. 6, closed-loop university hallways are typically 3 m wide with 90-degree turns, posing challenges for larger platforms attempting overtaking maneuvers in high-speed corners. This difficulty is further compounded in the 1 m wide puzzle mat circuits shown in Fig. 5, which are commonly used in classroom settings.

Fig. 6.

Hallway map: a 2D LIDAR map of the pepperdine hallways used for the experiments.

Fig. 5.

Overhead tracker: a logitech brio webcam was retrofitted with a machine vision lens to track the pocket racer platforms at 90 Hz.

Even at 1:28 scale, a limitation for our platform was that data collection via student demonstrations proved to be difficult for high speed cornering, which resulted in the throttle being fixed at a preset value. Other issues during hallway experiments included interference due to visual noise such as glare from sunlight and dynamic lighting in the evenings due to university sustainability policies. A robot vacuum was used to clear the carpet hallways from debris before testing, which also proved to be a challenge during testing. Finally, due to safety reasons, experiments had to be conducted in the evenings when students were not present in the hallways. These real world considerations suggested Pocket Racer’s advantage in classrooms despite the high speeds exhibited in the hallways all while highlighting the importance of a small form factor for accessible use.

The Pocket Racer platform introduced in this paper has a broad range of interdisciplinary applications in education¹⁵, especially within data science and engineering disciplines⁴⁹. As a pilot study to demonstrate Pocket Racer’s broad applications in interdisciplinary projects, we conducted an initial evaluation. At Seaver College, Pepperdine University, a robotics club was created during the summer of 2023 with n=6, where n stands for the number of students involved, in conjunction with the Physics and Computer Science department in an effort to promote inter-disciplinary education. Afterwards, the Pocket Racer platform was used as part of a project based curriculum for an applied machine learning class (n=4) within the Computer Science department.

Our platform was concurrently used at University of California, Los Angeles (UCLA) as a Special Projects Initiative (SPI) for the undergraduate IEEE club (n=38). This competition program is currently lead each year by two undergraduate IEEE club officers within the Electrical Engineering department. The students were able to obtain industry sponsorship to create a robotics curriculum as part of their extracurricular activities. Students from diverse fields of study with little or no experience in robotics helped design, build and prototype the Pocket Racer chassis.

The capacity of the Pocket Racer platform to exhibit overtaking maneuvers is determined by environmental conditions. In particular, Pocket Racer is sensitive to lighting conditions and to the presence of pedestrians or other unexpected obstacles in the camera field of view during both data collection and deployment. In addition, for head-to-head racing and overtaking, we recommend a minimum track width of 50 cm to allow safe side-by-side operation. Identifying the main limiting factors for operation are essential for safe and robust operation. Sufficient track width (< 50cm), scene cleanliness and lighting consistency are a few notable mentions. Pocket Racer’s broader applications involve larger racing competition events involving multiple schools⁷. Pocket Racer’s limitations for hallway use suggest future work may involve algorithmic development on image processing methods for robust deployment in mixed lighting conditions⁴⁰. Such development could have larger implications in the development of an off road platform for outdoor use²¹, furthering its potential as an versatile education platform.

Methods

Pocket Racer was designed to accessibly enable high speed overtaking for the purpose of AI education. Enabling accessible autonomous racing required the vehicle to be of low-cost with easy to procure parts for assembly. Consequently, the final platform design utilized only off-the shelf components easily obtainable online. A detailed bill of materials is provided in Table A1 of Appendix A.

As illustrated in the exploded view in Fig. 7, the assembly consists of a low-cost, $15 USD Raspberry Pi Zero 2W compute board was selected due to its ability to easily pair with 1. a Sony IMX708 monocular camera and 2. a generic Bluetooth gaming controller. As the camera was the sole sensory input to the deep neural network models, a 220 degree field-of-view (FOV) fisheye lens was positioned high above the chassis to allow for the detection of opposing vehicles during racing maneuvers. A Ready-to-Run (RTR) 1:28 scale, sub-$75 USD, remote control (RC) chassis was made to enable high speed autonomy on a low budget through several hardware modifications. First, the rear wheels utilized 14mm width rear wheels to increase stability with the use of soft compound tires to maximize traction on the carpet surface. Second, the center shaft enabling an All Wheel Drive (AWD) drivetrain was removed for a Rear Wheel Drive (RWD) conversion to enable longer run times by reducing weight and gear friction inefficiencies. The gear ratio was redesigned to prioritize top speed and the brushed motor was replaced with a brushless Direct Current (DC) motor (3500Kv) to avoid overheating during the high speed experiments. As a result, a toy speed measurement tool (Tamiya Speed Checker) was used to verify that the platform chassis was capable of achieving top speeds of 26 km/h.

Fig. 7.

Blowout diagram of the pocket racer platform, including the chassis, 3D-printed parts, powertrain, compute unit, and camera.

The Raspberry Pi Zero 2 W used a headless variant of the Operating System (OS) with SSH enabled to maximize computational resources and memory for model inference. The servo and Electronic Speed Controller (ESC) were controlled via two hardware Pulse Width Modulation (PWM) General Purpose Input Output (GPIO) pins on the compute module. The video files embedded with steering labels were saved on a micro SD card, then transferred to a GPU workstation for the training phase of the project. Once the model was trained, the micro SD card was reloaded with the trained model for deployment (Fig. 8).

Fig. 8.

Wiring diagram for the pocket racer platform.

Data collection and deployment

Testing experiments for data collection and algorithm deployment done on a reconfigurable, foam mat circuit (RCP Tracks) for both small and medium sized classrooms, as well as a 3m width, closed-loop hallway in the natural science building at Seaver College, Pepperdine University. These locations were selected to assess the platform’s readiness for educational use in indoor settings both in and out of the classroom. As a low cost two dimensional (2D) alternative to existing motion capture systems (MoCap), a 4k Logitech Brio Web camera was retrofitted with a machine vision, 130 degrees CS-mount lens for indoor tracking (Table A2). The system enabled 90 Hz state estimation with sub 5cm error on a circuit sized upwards of 11 feet by 16 feet.

In all field experiments, Pulse Width Modulation (PWM) signals are collected from the student user controlling the Pocket Racer platform on a Bluetooth gaming controller. During data collection, drop in battery State of Charge (SOC) resulted in a corresponding drop in voltage, which reduced the consistency of vehicle operation by causing fluctuations in motor speed. To mitigate this, a low-level Proportional-Integral-Derivative (PID) controller was utilized to regulate the throttle to a fixed target value, ensuring consistent vehicle speed throughout each run, allowing ease of use for undergraduate students. These student-operated platforms collected 2 hours worth of data (600k images @ 85 Hz) through 6 separate runs for 20 minutes each, roughly the time it took for the Lithium Polymer (LiPo) batteries to lose its State of Charge (SOC). Data was recorded within a .AVI format video stream with images consisting of 32 by 32 pixels. Vigilante effects due to fisheye lens resulted in unusable, black corner pixels and were consequently used to embed steering values rescaled from 0 to 255 for ease of data post processing.

Model training

During dataset pre-processing, FFMPEG codec tools were used to remove collisions, pedestrians and glare from sunlight within the recordings. Min-Max scaling converted the pixel values ranging from 0 to 255 back to 0 to 1. Model quantization was performed with TensorFlow Lite resulting in a 25% reduction in model size. Mixed precision training with Float16 was enabled to reduce training times. The hallway dataset consisted of 400k training targets of 2 stacked images with corresponding steering action labels. The validation and test sets consisted of 40k and 50k stacked images. The indoor circuit dataset consisted of 82k labels with 9k validation set and a 10k test set. Notably, the student-collected dataset was not shuffled to preserve temporal information at the expense of generalization capabilities.

Five convolutional layers with a kernel size of (3,3) and stride of (2,2) was used to optimize performance and latency for 2D CNNs and (3,3,3), (1,2,2) respectively for 3D CNNs with input shape as (stack, height, width, depth). These layers were used with batch normalization followed by 3 Multi-Layer Perceptron (MLP) layers of size 64 each. When using LRCNs, the TimeDistributed function in Tensorflow Keras was utilized with the convolution operations, followed by an LSTM layer of size 64 and 2 MLP layers of size 64 each respectively. For ViTs, we used mixed size image patches as tokens to a 2 layer attention encoder with 4 attention heads per layer, followed by MLP layers of size 64. For all model baselines, the batch size was set to 12000 and the Adam optimizer was used with learning rate set to 1e-4 via mean squared loss (MSE).

To replicate real world deployment constraints, all models were converted to the TensorFlow Lite format during evaluation. Input images were resized to pixels, and steering labels were normalized by subtracting the per-lap mode and dividing by 127.5. For temporal models (3D-CNN, 3D-ViT, and LRCN), consecutive frame pairs were used as input. To ensure consistency across datasets, ten laps were sampled uniformly at random from each dataset using a fixed seed of 42; when fewer than ten laps were available, all laps were retained for evaluation.

Evaluation metrics

Two distributional distance metrics quantify the similarity between predicted and ground truth steering distributions. Continuous steering values are discretized into bins spanning . Let and denote the normalized ground truth and predicted distributions, with cumulative distribution functions and .

The Hellinger Distance measures shape similarity between distributions. We use an unnormalized variant that omits the outer square root and factor, preserving the same monotonic ordering while providing a more sensitive scale. The Earth Mover’s Distance (EMD), or Wasserstein-1 distance, measures the minimum cost of transforming one distribution into another via their CDFs, capturing spatial shifts that the Hellinger Distance is insensitive to:

7

Pocket Racer best practices

To ensure reliable and repeatable experiments with the Pocket Racer platform, the following best practices should be observed prior to and during training:

1. Battery Safety. Lithium Polymer (LiPo) batteries must be charged inside a fireproof charger bag at all times to avoid fire hazards associated with battery swelling or puncture.

2. Front Differential Removal. The front differential of the chassis must be removed to convert the drivetrain from All Wheel Drive (AWD) to Rear Wheel Drive (RWD), reducing weight and gear friction inefficiencies for longer run times.

3. Suspension Maintenance. The front and rear wishbone suspension links must be kept free of debris such as human hair, which may introduce friction to rotational components and degrade performance during Pocket Racer experiments.

4. Steering Calibration. The distribution of steering values, including the midpoint neutral value and the full range of steering, must be recalibrated every time a crash occurs. Collisions may alter the original servo values due to the servo gears experiencing a mechanical skip in the servo gear tooth, resulting in steering bias.

5. ESC Arming. The Electronic Speed Controller (ESC) must be properly armed before operation to ensure that the throttle channel is active and responsive.

Recommendations for adopters and future research directions

As multi-agent testbeds become an integral component of makerspaces¹² and Project-Based Learning (PBL)⁸ educational frameworks, accessible platforms like Pocket Racer can serve as a physical medium through which students iteratively design, test, and compete with autonomous systems. Five scientific challenges are outlined as a practical starting point to guide future adoption for Pocket Racer.

1. Safe Multi-Agent Autonomous Driving. Safely collect edge cases in multi-agent driving scenarios to explore safety guarantees in multi-agent autonomy.

2. High School AI and Robotics Education Outreach. Support collaboration with high schools, where undergraduate students foster peer mentorship.

3. End-to-End Highway Driving. Evaluate end-to-end autonomy models for rapid prototyping in high-speed driving scenarios.

4. Team-Based Autonomous Racing. Explore social behaviors in multi-agent, team-based settings such as autonomous racing through game-theoretic frameworks.

5. Online and Lifelong Learning for End-to-End Autonomy via soft neuromorphic computing methods and approaches.

Guidelines and regulations

The experiments for Pocket Racer were conducted both at Seaver College, Pepperdine University and the Mechanical and Aerospace Engineering (MAE) department in the University of Caifornia, Los Angeles (UCLA). All experiments pertaining to data collection were performed in accordance to approved protocol, guidelines and regulations. All experimental protocols pertaining to lab safety were notified and approved by the Robotics and Mechanisms Laboratory (RoMeLa) at UCLA and by the Keck Institute of Data Science at Seaver College, Pepperdine University. Our work involved data collection by a robot using human driving for the purposes of the robot’s mechanical performance. All subjects (including undergraduate students) involved in data collection for the experiments submitted a written and signed informed consent form. (Note: Our human driving data is a control source that is not the variable of interest for the study and is completely anonymous. Thus, our work does not classify as research pertaining to human subjects in U.S. federal regulations under the Common Rule (45 CFR 46) and consequently does not need approval by the ethics committee or the Institutional Review Board of the respective academic institutions.)

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

J.C., V.E.T., K.C. and E.L. conducted the experiments, V.E.T., J.C. and S.W. analyzed the results. J.C., V.E.T., D.H. and D.K. edited the manuscript. All authors reviewed the manuscript.

Funding

This work was supported in part by the Keck Institute for Data Science at Seaver College, Pepperdine University and the Korea Institute for Advancement of Technology (KIAT) grant funded by the South Korean Government (Ministry of Trade, Industry and Energy, MOTIE) through the International Joint Technology Development Program for Industrial Technology International Cooperation (Project No. P0028050, Project Name: Development of a real-time integrated control SW platform for autonomous driving/collaborative robots using generative AI).

Data availability

We make available all of the data needed to evaluate the conclusions of our paper and provide an open source website. (Website: https://pocketracers.github.io/PocketRacerWebsite) Our open source code repository can be found in the file “PocketRacer.zip” and also on a Github repository. (Website: https://github.com/PocketRacers/PocketRacerRepo) The open source dataset is available upon request.

Declarations

Competing interests

The authors declare no competing interests.