The OpenCDA Open-source Ecosystem for Cooperative Driving Automation Research

I. Introduction

• Single-vehicle systems remain prone to occlusion and limited detection range, while mixed-autonomy traffic complicates safe decisions.These challenges are described as obstacles to large-scale real-world deployment.Source: §I · p. 1
• CDA uses machine-to-machine communication among vehicles, pedestrians, and infrastructure to support cooperation, extend perception range, and see through occlusion.The passage identifies CDA as a solution to challenges including occlusion and limited perception range.Source: §I · p. 1
• Academia often lacks real-world traffic data and computation resources, while industry pipelines may remain confidential and simulation setups are difficult to compare.These conditions motivate a shared ecosystem for CDA development and testing.Source: §I · p. 1
• OpenCDA grows from an earlier simulation tool into an ecosystem combining models, simulators, benchmark datasets, and scenario databases and generators.The ecosystem is organized around toolkits, scenarios, data, and communities.Source: §I · pp. 1–2
• Its toolchains include a simulation tool, deep-learning framework, and model libraries for prototyping and developing CDA systems and algorithms.The model libraries include state-of-the-art algorithms for automated-driving tasks.Source: §I · p. 2
• OpenCDA provides simulated and real-world datasets, benchmark scenarios for cooperative applications, and a plan to nurture a researcher and engineer community.The datasets support offline training for cooperative perception algorithms.Source: §I · p. 2

A. Simulators

• OpenCDA addresses a simulator gap by supporting full-stack CDA software development in simulation, unlike traffic-focused SUMO and Flow, which do not support regular driving components and CDA.CARLA is described as a popular open-source platform for automated-driving technology development.Source: A. Simulators · p. 2

B. Cooperative Driving Automation

• CDA applications include adaptive cruise control, platooning, cooperative merge, cooperative perception, and eco-driving at signalized intersections.CACC is longitudinal and communicates with the front vehicle; platooning can support multi-lane scenarios and cooperation with both the front vehicle and leader.Source: B. Cooperative Driving Automation · p. 2
• Cooperative perception shares visual cues through early, intermediate, or late fusion, and OPV2V introduced a large-scale simulated V2V perception dataset using OpenCDA.The passage also identifies OpenCOOD as an early cooperative detection framework.Source: B. Cooperative Driving Automation · p. 2
• OpenCDA, in partnership with FHWA’s CARMA Program, offers early-stage testing of customized algorithms for tasks including object-detection accuracy and overall traffic-safety pipeline assessment.CARMA develops vehicle and infrastructure software platforms and tools for full-scale software simulation and testing.Source: B. Cooperative Driving Automation · p. 2

III. The OpenCDA Ecosystem

• The ecosystem section describes major components and their interfaces as of January 2023.It explains how the components interact to form the ecosystem.Source: §III · p. 3

A. Development Toolkits

• OpenCDA’s development toolkits combine a simulation tool, the OpenCOOD deep-learning framework, and a model zoo of state-of-the-art algorithms.The toolkits support communities prototyping and developing CDA systems and algorithms.Source: A. Development Toolkits · p. 3

1) OpenCDA Simulation Tool:

• The simulation tool integrates CARLA for environment rendering, SUMO for background traffic, and DI-Engine for a reinforcement-learning environment.OpenCDA began as a co-simulation tool integrated with prototype CDA pipelines and regular driving components.Source: 1) OpenCDA Simulation Tool · p. 3
• OpenCDA extends regular automated-driving systems with modular CDA functions, including cooperative perception and platooning.The cooperative driving system processes sensor data through sensing, planning, and actuation layers; cooperative functions can replace default perception and planning when vehicles meet.Source: 1) OpenCDA Simulation Tool · p. 3
• The scenario manager assigns driving tasks and evaluates runs, while the data saver prepares sensor data for training and the loader replays simulated or qualifying real-world data.Real-world replay requires HD maps in OSM format and high-precision GPS data.Source: 1) OpenCDA Simulation Tool · p. 4

A. Development Toolkits

• OpenCOOD is a PyTorch framework for cooperative detection and segmentation that handles simultaneous multi-agent sensor data for the same scene.The paper presents it as a foundation for cooperative-perception research.Source: 2) OpenCOOD · p. 4

• YAML configuration and dataset classes let users adjust parameters and prepare multi-agent data for early, late, or intermediate fusion.Source: 2) OpenCOOD · p. 4
• The pipeline converts raw point clouds to voxels or bird’s-eye-view features, fuses multi-vehicle inputs, and formats model outputs for visualization and evaluation.Source: 2) OpenCOOD · p. 4, Figure 2 · p. 5
• Research teams have added models such as Where2Comm and When2Comm, plus support for CODD and multimodality.Source: 2) OpenCOOD · p. 5

• The library includes pretrained models usable within OpenCDA or on other platforms.Source: 3) Model Libraries · p. 5
• Its 3D LiDAR detection examples include VoxelNet, PointPillar, Pixor, and SECOND, while cooperative perception includes V2X-ViT, AttFuse, CoBEVT, and others.Source: 3) Model Libraries · p. 5
• Other library areas include localization, planning, platooning, and control, with examples such as Kalman filters, A* search, finite-state machines, PID, and MPC.Source: 3) Model Libraries · p. 5

• OpenCDA contains large-scale simulated and real-world datasets intended to support offline training for cooperative perception algorithms.Source: B. Data · p. 5
• The ecosystem released two large-scale simulated cooperative-perception datasets and collected real-world data using two L4 automated vehicles.Source: B. Data · p. 5
• Figure 3 shows data samples from different datasets in the OpenCDA ecosystem.Source: Figure 3 · p. 5

1) Simulated Dataset:

• OPV2V contains 11,464 frames of multi-vehicle LiDAR point clouds and RGB images, with 230K annotated 3D boxes for cooperative perception tasks.Source: 1) Simulated Dataset · p. 5
• V2XSet extends OPV2V with roadside units to support vehicle-to-infrastructure cooperation.Source: 1) Simulated Dataset · p. 6

2) Real-world Dataset:

• The real-world dataset includes multi-CAV LiDAR, RGB images, annotated 3D boxes, and HD maps; OpenCDA can create a CARLA digital twin from annotations and an xodr map file.Source: 2) Real-world Dataset · p. 6, Figure 4 · p. 6

C. Scenario Database

• OpenCDA provides rule-based standard scenarios and tools to manage or customize scenarios for testing cooperative-driving applications.Source: C. Scenario Database · p. 6

1) Rule-based Scenarios:

• V2XP-ASG perturbs collaborating agents’ poses and uses CARLA to render updated LiDAR observations, generating challenging scenes that can reduce perception performance; training in them can increase accuracy.Source: 2) Adversarial Scenarios · p. 6, 1) Rule-based Scenarios · p. 6

D. Community

• The Community of Practice provides feedback for the ecosystem roadmap and supports stakeholder engagement through webinars, panels, and workshops.Users may also create subsidiary tools for specific tasks.Source: 1) Ecosystem Growth · p. 6

1) Ecosystem Growth:

• Figure 5 compares ground-truth green boxes with predicted red boxes in a congested intersection and a highway entrance ramp in V2XSet.Source: Figure 5 · p. 7

D. Community

• The planned OpenCDA Automated Driving Challenge uses reproducible scenarios, pass ratio and collision rate, and full-stack and component tracks.Topics are tied to a roadmap developed through stakeholder engagement.Source: 1) Ecosystem Growth · pp. 6–7

1) Ecosystem Growth:

• The Community of Content Contributors helps develop and maintain OpenCDA, with monthly webinars and introductory materials intended to lower participation barriers.The core team plans start kits and tutorials for new users and contributors.Source: 2) Community · p. 7

3) Governance and Organization:

• OpenCDA currently uses a founder-led model, with UCLA's core team leading, and aims to move toward community-managed governance for sustainability and scalability.A foundation-backed model will also be explored; the choice considers flexibility, simplicity, scalability, and industry constraints.Source: 3) Governance and Organization · p. 7
• The software process assigns roadmap tasks to contributors, then routes pull requests through static and dynamic review and an online CI/CD build before merge.Contributors design, implement, test, and integrate functionality before requesting review.Source: 3) Governance and Organization · p. 7

A. Case Study 1: Cooperative 3D LiDAR Detection

• DiscoNet leads OPV2V by 1.0% on AP@0.5, while V2X-ViT beats DiscoNet by 7% on V2XSet and 4.2% on real-world data.Cooperative models outperform the No Fusion baseline; the table covers OPV2V and V2XSet at AP IoU thresholds of 0.5 and 0.7.Source: Table I · p. 8, A. Case Study 1: Cooperative 3D LiDAR… · p. 8
• The LiDAR study uses OpenCOOD on OPV2V, V2XSet, and a real-world dataset, with 14,000/2,000/4,000 frames for training, validation, and testing.A common data format supports OpenCOOD; evaluation uses vehicle detection and AP at IoU thresholds 0.5 and 0.7.Source: A. Case Study 1: Cooperative 3D LiDAR… · p. 8

B. Case Study 2: Cooperative Merge into Platoons

• A single ramp CAV requests to join a mainline platoon; the leader selects a merging position and directs members to create a gap before the acceleration lane ends.The scenario combines high-speed traffic managed by SUMO with CAVs controlled by cooperative driving systems.Source: B. Case Study 2: Cooperative Merge… · p. 8
• Evaluation covers safety, platoon stability, and joining efficiency, using TTC and hazard frequency, time-gap and acceleration measures, and maneuver duration and acceleration.The TTC warning threshold is 2.5 seconds.Source: B. Case Study 2: Cooperative Merge… · pp. 8–9
• The heuristic selects the platoon vehicle with the shortest Euclidean distance as the frontal vehicle, whereas GFS uses fuzzy logic to choose the merging position.GFS also considers platoon-member speeds and surrounding human-driven vehicles.Source: B. Case Study 2: Cooperative Merge… · p. 9
• Both merging algorithms achieve good safety and stability, while the Genetic Fuzzy System performs favorably against the heuristic method in efficiency.The desired platoon time gap is 0.6 seconds.Source: B. Case Study 2: Cooperative Merge… · p. 9, Table II · p. 9

C. Case Study 3: Camera-based Cooperative Perception

• The study uses four cameras per connected vehicle and combines inferred 3D information from multi-view images to predict a BEV segmentation mask.Because OPV2V lacks map information, driving-log replay generates HDMap ground truth.Source: C. Case Study 3: Camera-based… · p. 10
• Testing evaluates IoU within a 100 m × 100 m area around a fixed ego vehicle at 39 cm map resolution, segmenting dynamic objects and map topology.Source: C. Case Study 3: Camera-based… · p. 10
• CoBEVT achieves the best IoU for all classes, beating the second-best method by 7.5% for vehicle, 2.3% for drivable area, and 7.2% for lane.All cooperative methods outperform No Fusion; the authors report that collaboration overcomes most occlusions and perceives distant objects accurately.Source: C. Case Study 3: Camera-based… · p. 10

D. Case Study 4: Adversarial Scene Generation

• V2XP-ASG generated challenging scenes with high AP drops across V2X fusion methods, and qualitative examples show vehicle misdetections.The comparison used generated scenes against original normal scenes; Fig. 10 shows the qualitative results.Source: D. Case Study 4: Adversarial Scene… · p. 10, Figure 10 · p. 11
• V2XP-ASG searches for adversarial collaborators using intermediate-fusion attention weights, then uses black-box optimization to find perturbations linked to inferior performance.It uses historical perturbation and loss pairs, and filters perturbations that could cause collisions or place agents out of range.Source: D. Case Study 4: Adversarial Scene… · p. 10
• The evaluation used 331 augmented OPV2V scenes in CARLA, with the ego fixed, zero LiDAR noise, and agents evaluated within x, y ∈ [−48,48] m.The setup isolates performance drops due to scene configurations, according to the passage.Source: D. Case Study 4: Adversarial Scene… · p. 10

V. Conclusions

• OpenCDA integrates a model zoo, driving simulators at varied resolutions, benchmark datasets, and a scenario database and generator for CDA research.The conclusion says case studies demonstrate the ecosystem’s capability to facilitate CDA research from different perspectives.Source: §V · pp. 10–11
• Planned additions include a human interaction layer, end-to-end learning, a larger model zoo, real-system interfaces, and a ROS simulation version.The planned ROS version is intended for hardware-in-the-loop testing and direct deployment of cooperative-driving components on real vehicles.Source: §V · p. 11
• The authors say OpenCDA’s continued development will rely on community contributions as it moves from founder-led to community-managed.Source: §V · p. 11