一句话：提出执行适配器 PCLA-Bridge 与可复用性分类法，使 CARLA 训练的驾驶智能体无需重训即可在真实数据上运行。
论文关注 CARLA Leaderboard 上的高性能自动驾驶智能体通常受限于模拟器专属 API、传感器格式与坐标系，难以直接部署到真实日志或实车。作者在 PCLA 框架上提出 PCLA-Bridge：一个无需 CARLA 服务器即可从任意传感器流驱动智能体的执行适配器，并给出可复用性分类法以刻画影响可移植性的架构约束。评估通过离线回放进行，覆盖域内 CARLA 日志与真实 nuScenes 数据集。未修改的 bridge 成功用 nuScenes 物理数据驱动 CARLA 训练的智能体，生成准确的转向指令，并通过观察到的油门谨慎行为量化了 sim-to-real 域差距。
While many high-performing autonomous driving agents (ADAs) are published via the CARLA Leaderboard, they remain rigidly bound to the simulator's specific APIs, sensor formats, and coordinate frames. This work investigates whether these simulation-trained models can be deployed outside their native environments, such as on real-world logs or physical vehicles, without retraining. Building upon the PCLA framework, we introduce two contributions: (1) PCLA-Bridge, an execution adapter that drives agents from arbitrary sensor streams without requiring an active CARLA server; and (2) a reusability taxonomy defining the architectural constraints that govern agent portability. We evaluate PCLA-Bridge via offline replay on both in-domain CARLA logs and real-world nuScenes datasets. The unmodified bridge successfully drives a CARLA-trained agent using physical nuScenes data, producing accurate steering commands, while effectively quantifying the sim-to-real domain gap through observed throttle caution.