PlannerForge: LLM Agents for Scenario-Based Testing of Motion Planners in Autonomous Driving
PlannerForge turns the fragmented motion-planner testing pipeline into one LLM-agent workflow.
The paper says the framework spans scenario generation, selection, modification, planner testing, assessment, enhancement, and benchmarking. In tests across 10 off-the-shelf LLMs and five prompt conditions, best task scores ran from 0.88 to 1.00. Open-source 20-35B models matched commercial APIs on most tasks, with Qwen3.6:35B matching on three of five. At 400 scenarios, cost-tuning raised planner success from 50.4% to 70.2% and cut collisions from 19.0% to 8.4%, without domain-specific fine-tuning. Source: HF Daily Papers' note.
The paper says the framework spans scenario generation, selection, modification, planner testing, assessment, enhancement, and benchmarking. In tests across 10 off-the-shelf LLMs and five prompt conditions, best task scores ran from 0.88 to 1.00. Open-source 20-35B models matched commercial APIs on most tasks, with Qwen3.6:35B matching on three of five. At 400 scenarios, cost-tuning raised planner success from 50.4% to 70.2% and cut collisions from 19.0% to 8.4%, without domain-specific fine-tuning. Source: HF Daily Papers' note.
score 4