Workshops Archive

HPC-CARLA: Closed-Loop Evaluation of Autonomous Driving Agents as a Resilient, Schedulable HPC Workload


Workshop: 21st Workshop on Workflows in Support of Large-Scale Science (WORKS26)

Authors: Jey Kang (Gwangju Institute of Science and Technology), Seungbum Jung (Saint Louis University), Kangryeol Kim (Gwangju Institute of Science and Technology), Cory Gardner (Saint Louis University), Jae-Seung Yeom (Lawrence Livermore National Laboratory), Tae-Hyuk Ahn (Saint Louis University), and JongWon Kim (Gwangju Institute of Science and Technology)

Abstract: Closed-loop evaluation of autonomous-driving policies spans many conditions, but CARLA servers on shared GPU clusters often fail after producing valid episodes. We present HPC-CARLA, which runs these campaigns through a multi-node queue and reuses one simulator service per GPU. It restarts failed services and harvests episode checkpoints independently of job state. Its declarative pipeline framework brings six policies under one audited execution model. Across five agents on A100, HPC-CARLA retained 13,059 valid route evaluations; 94.3% came from jobs that ultimately failed. Three L40S campaigns tested another GPU and driver stack, although score agreement depended on the agent. As A100 coverage expanded, the relationship between an offline difficulty score and driving performance reversed, so the revised dispatcher gives per-agent coverage requirements precedence over that scalar. Together, the campaigns show why workflow state must include simulator health, episode progress, and condition coverage.


Back to 21st Workshop on Workflows in Support of Large-Scale Science (WORKS26) Archive Listing

Back to Full Workshop Archive Listing