PCLA: Pretrained CARLA Leaderboard Agents
A framework for deploying, testing and evaluating pretrained autonomous-driving agents from the CARLA Leaderboard on a vehicle in the CARLA simulator, decoupled from the Leaderboard's own codebase.
PCLA (Pretrained CARLA Leaderboard Agents) lets a developer deploy, switch between and evaluate 41 pretrained autonomous driving agents and 27 additional training seeds from 18 major autonomous-driving research projects, including SimLingo, TransfuserV6, Roach, Interfuser, NEAT and LAV, on a vehicle in the CARLA simulator, without depending on the Leaderboard's own restrictive codebase. An added Autoware bridge lets it also drive Autoware, an open-source, HD-map-based modular autonomous-driving stack, rather than only end-to-end neural-network agents.
Switching between agents and training seeds needs no change to the CARLA version or programming environment, and the framework is compatible with the latest CARLA release independent of the Leaderboard's own pinned version. It can run multiple vehicles with different agents at once, though this needs substantial graphics memory, and a separate PCLA bridge lets an agent's computed driving action be sent to any external sensor source or use case without needing the CARLA simulator itself.
It was tested on Linux Ubuntu 22 with CARLA 0.9.16, and the accompanying paper appeared at the Foundations of Software Engineering conference.