Source-linked AI summary
PCLA-Bridge: A Framework for Cross-Domain Deployment of CARLA Agents
Masoud Jamshidiyan Tehrani, Jinhan Kim, Paolo Tonella
TL;DR
CARLA agents are coupled to their original execution harness, limiting reuse on other data sources and control targets. PCLA-Bridge runs unmodified agents without an active CARLA server, handling agent isolation and input adaptation. The paper reports offline replay on CARLA and nuScenes data and unmodified-agent execution across three foreign simulators.
Problem
CARLA agents depend on their original simulator’s sensors, routing, timing, and coordinate conventions, making reuse on other sources or platforms require manual re-engineering.
Source: §I · p. 1Method
PCLA-Bridge isolates unmodified agents and handles sensor marshalling, route setup, and synthetic timing while external sources provide inputs and targets consume outputs.
Source: §I · p. 1, §II · p. 1Results
Agents generated control outputs from CARLA recordings and nuScenes logs, and portable agents executed unmodified in closed loop across MetaDrive, AWSIM, and AlpaSim.
Source: §V · p. 3Takeaways & Limitations
PCLA-Bridge makes evaluation of CARLA agents on external data and simulators a lightweight integration task rather than agent re-engineering.
Source: §VI · p. 3, §VII · p. 3The approach focuses on sensor-only agents; privileged agents requiring host-simulator ground truth cannot be driven by empirical sensors.
Source: §II · p. 1
Abstract
from arXiv · showhide
Most high-performing end-to-end autonomous driving agents (ADAs) are developed for the CARLA Leaderboard, tightly coupling them to CARLA's APIs, sensor layouts, coordinate systems, and internal clock. Reusing these models on real datasets, alternative simulators, or physical platforms necessitates substantial manual re-engineering. We present PCLA-Bridge, a deployable tool that runs unmodified CARLA Leaderboard agents from any sensor source to any control destination without an active CARLA server. With PCLA-Bridge handling dependency isolation, tensor marshalling, routing, and synthetic timing, integrators only need to provide a minimal interface. We validate the tool across two operational paradigms: offline replay on in-domain CARLA recordings and real-world nuScenes logs, and live closed-loop control across three diverse simulation engines: MetaDrive, AWSIM, and AlpaSim. We analyze the lightweight integration requirements for each platform, highlighting output translation (such as converting agent actions into target trajectory formats) and coordinate-frame transformations, showing that integration requires minimal glue code rather than structural changes. All eight evaluated agents execute across all three simulators without modification, confirming the reliability of PCLA-Bridge and establishing a foundation for systematic sim-to-real, cross-sim, and security testing.
I. INTRODUCTION
CARLA agents are tightly coupled to their original execution harness, making reuse across datasets, simulators, and physical platforms labor-intensive. PCLA-Bridge aims to let researchers evaluate them on new inputs without modifying or rewriting the agents.
- CARLA agents depend on native sensor tensors, routing, temporal synchronization, and coordinate conventions, so reuse across environments requires manual re-engineering.Source: §I · p. 1
- PCLA-Bridge runs unmodified agents without a CARLA server, letting any data source drive an agent and any control target receive its outputs.Unlike the earlier offline-replay concept, this paper describes live closed-loop operation and practical integration requirements.Source: §I · p. 1
- The intended users are autonomous-driving researchers and software-engineering testers evaluating agents on inputs unavailable from their native simulator, including logs, other simulators, perturbations, or vehicles.Source: §I · p. 1
II. PCLA-BRIDGE
PCLA-Bridge treats agent execution as simulator-agnostic and supplies the isolation and data handling needed to connect agents to external sources and targets. Portability still depends on the agent’s input and output architecture.
- An agent’s per-frame execution is treated as simulator-agnostic: it receives sensor data and returns control values.Source: §II · p. 1
- The bridge loads each agent in an isolated environment and marshals inputs by resizing cameras, zero-filling absent modalities, setting GPS-waypoint routes, and managing synthetic time.The vendored CARLA module resolves data types; it does not require an active server connection.Source: §II · p. 1
- Sensor-only agents can be ported to new environments, but privileged agents need host-simulator ground truth that empirical sensors cannot provide.Source: §II · p. 1
- Systems without a monolithic sensor-in, control-out architecture, such as Autoware, cannot be ported using this approach.Source: §II · p. 1
III. RELATED WORK
Related work addresses sim-to-real transfer and the limitations of current autonomous-driving evaluations. PCLA improved agent testability but remained limited to CARLA.
- Sim-to-real approaches include generative perception bridging [18], Gaussian splatting for sensor rendering, and policy adaptation [20], while top CARLA-scoring agents have transferred poorly to reality.Source: §III · p. 2
- Open-loop nuScenes benchmarks can suffer data leakage and favor straight paths, while NAVSIM evaluates predicted trajectories on recorded data as a compromise with closed-loop testing overhead.Source: §III · p. 2
- PCLA improved agent testability but remained constrained to CARLA.Source: §III · p. 2
IV. BRIDGING TO A NEW SIMULATION
PCLA-Bridge was integrated with three simulators spanning different rendering methods, scene origins, software stacks, and sensing modalities. Each integration used lightweight adapters for control translation, frame calibration, and sensor mapping.
IV. BRIDGING TO A NEW SIMULATION
- The evaluation spans MetaDrive, AWSIM, and AlpaSim, from procedural scenes and a photorealistic digital twin to neural reconstructions of physical driving logs.The engines also differ in software stack and sensing; the progression is described as increasing perceptual realism toward physical deployment.Source: §IV · p. 2
- Each simulator integration required only three lightweight components—control-output translation, frame calibration, and sensor mapping—with no further modifications.Source: §IV · p. 2
- Adapters translate throttle, steer, and brake into each simulator’s control format; AlpaSim instead receives a recomputed short-horizon trajectory expressing instantaneous agent intent.AlpaSim’s adapter scales steering to wheel angle, maps throttle and brake to target speed, and generates waypoints for its internal path tracker.Source: §IV · p. 2
- Frame calibration uses constants for differences such as GPS-axis inversion, compass offsets, and steering signs, calibrated against a reference drive.An AlpaSim ego-pose sign discrepancy was corrected to make ego pose consistent with the synthetic GPS signal.Source: §IV · p. 2
- When AlpaSim omits LiDAR, PCLA-Bridge zero-fills the missing channels so camera-LiDAR agents can operate in a camera-only configuration.Source: §IV · p. 2
- The AlpaSim integration ran the agent and GPU renderer on separate machines, communicating through a network socket over an SSH tunnel without structural code changes.Source: §IV · p. 2
- Table II reports on-route distance in meters per agent across three simulators and two live scenarios.Source: Table II · p. 3
V. USAGE DEMONSTRATION
The evaluation tests eight sensor-only agents in offline replay and live closed-loop settings. Agents execute unmodified across three simulators, while real-world nuScenes replay shows coherent steering alongside near-zero throttle.
Evaluation setup
- The evaluation covers eight top-tier sensor-only agents: four camera-only models and four camera+LiDAR models.The authors state that this selection lets observed performance differences isolate portability rather than underlying agent deficiencies.Source: §V · p. 3
Offline replay (open-loop).
- Offline replay runs every agent without a CARLA server and produces route-tracking control on CARLA recordings and a left turn on nuScenes imagery.On unfamiliar real-world imagery, throttle is near zero, reflecting an overly cautious policy.Source: §V · p. 3
Live sim-to-sim (closed-loop).
- Every portable agent executes unmodified in closed-loop control across MetaDrive, AWSIM, and AlpaSim using a lightweight adapter.Route coverage is summarized for baseline scene A and more challenging scene B in each simulator.Source: §V · p. 3
VI. WHAT PCLA-BRIDGE ENABLES
PCLA-Bridge shifts several evaluation tasks from re-engineering to one-line integrations. The proposed uses span sim-to-real measurement, security testing, cross-simulator diagnosis, offline regression, and standardized benchmarking.
Quantifying the sim-to-real gap
- Running CARLA-trained policies on real-world datasets can quantify sim-to-real degradation, including near-zero throttle under empirical camera imagery.The observed hesitation is offered as an example of behavioral degradation.Source: §VI · p. 3
Security and adversarial transferability
- Physical attack sensor streams can test whether vulnerabilities found inside CARLA persist under real-world conditions.Source: §VI · p. 3
Cross-simulator fault localization
- Keeping the agent architecture fixed while changing simulators can help isolate driving violations to model defects or host-specific physics and rendering anomalies.This is presented as cross-simulator fault localization.Source: §VI · p. 3
Scalable offline regression testing
- Recorded sensor trajectories can support scalable offline regression over thousands of kilometers without the steep compute overhead of a GPU-bound Unreal Engine instance.Source: §VI · p. 3
Benchmark standardization
- Unified execution standards provide a consistent testing ground, while scene-dependent Table II variance shows no single agent dominates across all engines.Source: §VI · p. 3
VII. CONCLUSION AND FUTURE WORK
PCLA-Bridge makes sensor-only CARLA-trained agents reusable without an active CARLA simulation, producing control from recorded data and external simulators. Physical-vehicle validation, portability metrics, and adversarial-transfer studies remain planned future work.
Conclusion
- PCLA-Bridge reliably generates control outputs from recorded CARLA and nuScenes data and supports unmodified agents in closed-loop control across three foreign simulators despite domain-shift performance variations.Source: §VII · p. 3
Future work
- Future work plans physical-vehicle validation, behavioral fidelity and portability metrics, and adversarial transferability studies across sim-to-sim and sim-to-real boundaries.The stated aim of the studies is to uncover intrinsic agent vulnerabilities without simulator biases.Source: §VII · p. 3