Source-linked AI summary
F1/10: An Open-Source Autonomous Cyber-Physical Platform
Matthew O'Kelly, Varundev Sukhil, Houssam Abbas, Jack Harkins, Chris Kao, Yash Vardhan Pant, Rahul Mangharam, Dipshil Agarwal, Madhur Behl, Paolo Burgio, Marko Bertogna
TL;DR
Real-vehicle AV testing is costly and hazardous, and affordable, open-source integrated testbeds suitable for typical indoor CPS labs are lacking. The paper presents F1/10, an open-source 1/10-scale autonomous racing platform using ROS, with open-source mechanical designs and AV software libraries for perception, planning, control, and coordinated autonomy research. The platform provides realistic dynamics for safe, cost-effective algorithm testing, and the paper demonstrates representative research examples including obstacle avoidance, lane keeping, and end-to-end driving.
Problem
Real-vehicle AV testing is costly and hazardous, and affordable, open-source integrated testbeds suitable for typical indoor CPS labs are lacking.
Source: §1 · p. 1Method
The paper presents F1/10, an open-source 1/10-scale autonomous racing platform using ROS, with open-source mechanical designs and AV software libraries for perception, planning, control, and coordinated autonomy research.
Source: §1 · pp. 1–2Results
The platform provides realistic dynamics for safe, cost-effective algorithm testing, and the paper demonstrates representative research examples including obstacle avoidance, lane keeping, and end-to-end driving.
Source: §8 · p. 9Takeaways & Limitations
F1/10 takes care of testbed assembly so researchers and students can focus on autonomous-systems research and learning goals.
Source: §1 · p. 2
Abstract
from arXiv · showhide
In 2005 DARPA labeled the realization of viable autonomous vehicles (AVs) a grand challenge; a short time later the idea became a moonshot that could change the automotive industry. Today, the question of safety stands between reality and solved. Given the right platform the CPS community is poised to offer unique insights. However, testing the limits of safety and performance on real vehicles is costly and hazardous. The use of such vehicles is also outside the reach of most researchers and students. In this paper, we present F1/10: an open-source, affordable, and high-performance 1/10 scale autonomous vehicle testbed. The F1/10 testbed carries a full suite of sensors, perception, planning, control, and networking software stacks that are similar to full scale solutions. We demonstrate key examples of the research enabled by the F1/10 testbed, and how the platform can be used to augment research and education in autonomous systems, making autonomy more accessible.
1 INTRODUCTION
F1/10 addresses the scarcity of accessible, realistic AV testbeds with an open-source 1/10-scale platform for research and education. Its modular hardware and software, research examples, and community activities make it a shared instrument for autonomous-systems work.
1 INTRODUCTION
- Isolated simulation, proprietary hardware, and the need to build testbeds from scratch constrain realistic, reproducible CPS and AV research.The paper notes that access to manufacturers’ documentation may also be restricted, preventing release of resulting hardware and software.Source: §1 · p. 1
- F1/10 is an open-source, affordable 1/10-scale autonomous-vehicle testbed intended to serve as a common research and education platform.The paper reports that about 20+ institutions use the testbed.Source: §1 · p. 1
- Open-source hardware and ROS-based software combine modular design with realistic vehicle dynamics and high-speed capability.The platform is contrasted with TurtleBot2 and Jackal UGV, which the paper says lack realistic Ackermann steering and high-speed travel.Source: §1 · p. 2
- The paper presents more than a dozen research examples, including testing AV algorithms and software pipelines with realistic dynamics on a physical, affordable testbed.The open-source capabilities include mechanical designs and assembly kits, AV libraries, a simulator and virtual track, competitions, educational programs, course material, and datasets.Source: §1 · p. 2
- The contributions also include a case study moving from 1/10-scale cars toward full-scale autonomous vehicles and an overview of competitions at CPS and embedded-systems venues.The competitions have been held at premier venues over the last three years.Source: §1 · p. 2
2 F1/10 TESTBED
F1/10 is designed to reproduce full-scale vehicle dynamics while remaining modular and self-sustaining. Its ROS-based architecture connects sensing, planning, and control, and its build combines an accessible race-car chassis with onboard computation and sensors.
2 F1/10 TESTBED
- The platform is designed to capture full-scale autonomous-car dynamics, support modular upgrades and repairs, and operate without external localization.The stated requirements include self-sufficient power, computation, and sensors, without VICON cameras.Source: §2 · p. 2
2.1 System Architecture
- The ROS architecture links perception sensors to mapping and path planning, then converts the resulting trajectory into steering and acceleration commands.The described sensors include scanning LiDARs, monocular and stereo cameras, and inertial sensors.Source: §2.1 · p. 2
2.2 F1/10 Build
- The build combines a Traxxas 1/10 chassis with a custom laser-cut ABS electronics plate and a Jetson TX2 running the ROS stack.The default sensors include a monocular webcam, ZED depth camera, Hokuyo 10LX scanning LiDAR, and MPU-9050 IMU; the software stack is available as a flashable image.Source: §2.2 · p. 3
- A power board stabilizes battery voltage and simplifies peripheral wiring, while onboard VESC odometry and wireless ROS networking support vehicle operation.The networking setup supports remote commands and telemetry, as well as communication between vehicles using different ROS master nodes.Source: §2.2 · p. 3
3 RESEARCH: PLANNING AND CONTROL
F1/10 supports research across obstacle avoidance, end-to-end driving, path planning, model predictive control, and vehicle-to-vehicle collaboration. Examples range from a learned planner approximation with high trajectory-generation throughput to experiments in connected-vehicle scenarios.
3.1 Obstacle avoidance
- Follow the Gap steers toward the largest LIDAR gap while fusing that heading with the vehicle’s goal.The method builds a gap array, finds the maximum gap, and calculates a steering value toward its center.Source: §3.1 · p. 3
3.2 End-to-end driving
- An end-to-end driving system combines LSTM and CNN units in a recurrent network, predicting steering angle with an NRMSE of 0.14.The platform also supports teleoperated data collection; the released dataset contains over 40,000 labeled images.Source: §3.2 · pp. 3–4
3.3 Global & local approaches to path planning
- Global planning can use track maps and offline paths, while local planning tracks the path with real-time localization and control.The software includes pure pursuit, waypoint, and path-visualization tools.Source: §3.3 · p. 4
3.4 Model Predictive Control
- The MPC planner optimizes cubic-spline trajectories between the vehicle’s current state and a goal pose.The spline parameters encode curve length and curvature at equispaced knots; the learned approximation maps target states to trajectory parameters.Source: §3.4 · p. 5
- 0.1% worst-case test error and over 428,000 trajectories per second on 145,824 test samples for the learned MPC approximation.The approximation uses a radial basis function network.Source: §3.4 · p. 5
3.5 Vehicle-to-Vehicle Communication, Cooperation, and Behavioral Planning
- F1/10 supports V2V roundabout experiments where vehicles exchange relative positions and whether their preferred next action is safe.The platform provides an intermediate step between simulation and real-world testing for studying sensor noise, wireless degradation, and actuation error.Source: §3.5 · p. 5
4 RESEARCH: PERCEPTION
F1/10 supports perception research in the context of driving tasks, letting researchers compare methods on a uniform platform rather than only in isolation. Its software and sensor capabilities cover SLAM, computer vision, and lane keeping.
4 RESEARCH: PERCEPTION
- F1/10 lets researchers compare perception methods by their whole-system capabilities within operating regimes and specific driving tasks.The car’s planning and state-estimation capabilities provide a common platform for those comparisons.Source: §4 · p. 5
4.1 Simultaneous Localization and Mapping
- The platform provides a LIDAR-based Google Cartographer interface and empirically tested local and global settings for SLAM.A GPU raymarching localization package can support research on driving at the limits of control without motion capture.Source: §4.1 · pp. 5–6
4.2 Computer Vision
- F1/10’s software includes CUDA-enabled deep-learning frameworks and example networks for segmentation, object detection, and optical flow.The included efficient variants are described as running at greater than 10 FPS on the TX2; the platform can support studies of how DNN vision errors affect vehicle-level performance.Source: §4.2 · p. 6
4.3 Lane keep assist
- A camera-based lane-keeping pipeline detects lane markings, calculates lane center, and uses a PID controller to adjust heading.The image processing includes color filtering, edge detection, Gaussian filtering, and a Hough transform.Source: §4.3 · p. 6
5 RESEARCH: SYSTEMS, SIMULATION, AND VERIFICATION
F1/10 supports systems research spanning simulation, real-time scheduling, verification, monitoring, and traffic experiments. The platform allows these tools and methods to be examined with vehicle hardware and workloads, while simulation offers a controlled step before real-world testing.
5.1 Gazebo Racing Simulators
- Gazebo simulation supports controlled algorithm testing before real-world deployment and lets resource-constrained users complete course labs with the software stack.The package includes four tracks with real-world counterparts and allows custom environments.Source: §5.1 · pp. 6–7, Figure 5 · p. 7
5.2 Real-time Systems Research
- The F1/10 stack is suited to testing real-time scheduling and task partitioning because timing, localization, perception, and control delays bear on vehicle stability and safety.The platform also supports research on CPU/GPU allocation and execution frequencies that trade computation time, power, and output quality while maintaining control performance.Source: §5.2 · p. 7
5.3 Monitoring, Testing, & Verification
- F1/10 supports real-time testing of reachability software running alongside other workloads on target hardware.Reachability analysis bounds future states despite uncertainty in the current state, helping a car avoid unsafe plans.Source: §5.3 · pp. 7–8
- The platform can test generated runtime monitors, including their efficiency and deployment across multiple processors.F1/10 competitions could test how quickly monitors can be created and deployed as conditions change between practice laps and the next round.Source: §5.3 · p. 8
- Reduced-scale F1/10 fleets can provide traffic for navigation tests and generate more data cheaply than a full-scale fleet for evaluating statistical safety claims.A learned traffic model can drive the fleet in a test setup.Source: §5.3 · p. 8
6 FROM F1/10 TO FULL-SCALE AVS
The F1/10 open-source AV stack provided a starting point for a full-scale autonomous-car prototype at the University of Modena. Its routines supported rapid development and iterative component work, while the deployed car later replaced ROS with a stack offering stronger real-time guarantees.
6 FROM F1/10 TO FULL-SCALE AVS
- The full-scale prototype extended F1/10’s perception-planning-control stack with drivers and routines for six cameras, a 3D lidar, and differential GPS.Its first autopilot used pure pursuit for trajectory tracking and model predictive control for trajectory generation; Gazebo and Unity were used for pre-deployment testing.Source: §6 · p. 8
- F1/10 routines were instrumental in deploying a working full-scale AV prototype in limited time and provided a sandbox for iterating on its components.The prototype could navigate roundabouts and line-marked paths, avoid detected obstacles, stop at traffic signals, and park in a user-defined slot.Source: §6 · p. 8
- The prototype included lane keeping, object detection, obstacle avoidance, SLAM, camera-lidar fusion, and V2I communication, but its current version replaced the ROS-based meta-operating system for stronger real-time guarantees.Figure 6 shows the F1/10 perception, planning, and control pipeline applied in designing full-scale autonomous racecars at UNIMORE, Italy.Source: §6 · p. 8, Figure 6 · p. 8
7 F1/10 EDUCATION AND COMPEITIONS
F1/10 supports courses that teach autonomous driving across the vehicle stack, and competitions that invite teams to build and race cars using a shared baseline. The education and competition programs have reached students across several levels and teams from more than 12 universities.
7 F1/10 EDUCATION AND COMPEITIONS
- F1/10-based courses teach the autonomous-driving stack from assembling components and sensors to programming the car, and have been offered at three universities and taught using open materials at two others.UVA online videos had been viewed thousands of times; the courses serve graduate and advanced undergraduate students from several engineering and computing backgrounds.Source: §7 · p. 8
- The project-based course guides teams from building and manually controlling a car to navigating static obstacles, covering LiDAR processing, localization, mapping, and planning.It also includes readings and discussions on responsibility, moral agency, and ethics for autonomous robots, and culminates in a team “battle of algorithms” race.Source: §7 · p. 9
7.1 Te F1/10 Competition
- Competition teams build the same baseline car, then choose algorithms to maximize loop speed and laps completed within a fixed duration.The international competition had been held at ES-Week in 2016 and 2018 and at CPS Week in 2018; future editions were planned to feature car-vs-car racing.Source: §7.1 · p. 9
- Teams from more than 12 universities had participated in the F1/10 competition.Figure 7 presents outreach through online courses and MOOCs, competitions, summer schools, and hackathons.Source: §7.1 · p. 9, Figure 7 · p. 9
8 CONCLUSION AND DISCUSSION
The paper concludes that F1/10 is an open-source, modular 1/10-scale AV testbed that supports safe, cost-effective real-world testing. It describes research, education, and community uses of the platform.
8 CONCLUSION AND DISCUSSION
- F1/10 provides a modular, open-source 1/10-scale AV testbed with freely available build, driving, and racing instructions and a default configuration with sensors and an onboard GPU.Its chassis provides realistic dynamics for safe, cost-effective algorithm testing at 1/10 scale.Source: §8 · p. 9
- The paper presents obstacle avoidance, lane keeping, and end-to-end autonomous driving as research examples enabled by F1/10.It also reports that dozens of research groups built cars using the project’s instructions and videos.Source: §8 · p. 9
- F1/10 is presented as a vehicle for teaching the interlocking concerns of performance, control, and safety, and as a meeting point for researchers to develop ideas through prototyping and real-world testing.The authors state that the range of possibilities grows as the community continues to grow.Source: §8 · pp. 9–10