Research map / Large language models / Agents, code and math

Generating Traffic Scenarios via In-Context Learning to Learn Better Motion Planner

AAAI 2025 · Large language models › Agents, code and math · approach: LLM-driven embodied and driving agents · 5 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AutoSceneGen, which uses GPT-4 with in-context learning to turn text scenario descriptions into CARLA simulator scripts, and shows that the resulting synthetic trajectory data improves existing motion/trajectory prediction models.

Problem

Motion planners need diverse, safety-critical traffic scenarios that are rare in real-world datasets, and manually scripting such scenarios in simulators is costly.

Approach

Represent traffic scenarios as simulator scripts and use an LLM with in-context exemplars (plus a filter and validator for simulator API compatibility) to translate user text descriptions into executable CARLA configurations, with no training or fine-tuning of the generator.

Key results

• Trained TrafficPredict on generated data alone reports ADE-pedestrian 0.020 and FDE-pedestrian 0.037 versus 0.091 and 0.150 on the ApolloScapes-only baseline.
• On TraPHic, generated data alone gives TAE 0.27 versus 5.10 for the reproduced ApolloScapes baseline.
• Pishgu on NGSIM with generated data alone is much worse (ADE 7.98 vs 0.88), but combining NGSIM with generated data lowers ADE from 0.88 to 0.84.

Measured effects reported in the paper

Surprising findings

• Generated-only data was worse than real NGSIM for Pishgu on NGSIM (ADE 7.98 vs 0.88), while the combined training set improved results.
• Generated data alone outperformed real ApolloScapes on several TrafficPredict and TraPHic metrics despite being produced from only 125 text descriptions.

Limitations

• Performance depends on the LLM's in-context learning ability and the simulator's capabilities; hard corner cases can be difficult to generate.
• Pedestrian simulation in CARLA is limited, so no improvement was seen on VIRAT/ActEV.
• Explainability of generated scenario logic is limited to the descriptions and ICL examples; regulatory and traffic-law variation is not handled.

Experimental setup

Datasets: ApolloScapes TRAF NGSIM ETH/UCY VIRAT/ActEV

Benchmarks: ApolloScapes TRAF NGSIM ETH/UCY VIRAT/ActEV

Metrics: ADE FDE TAE TFE

Baselines: TrafficPredict TraPHic Pishgu TRAF

Builds on: TrafficPredict TraPHic Pishgu MSRL CARLA

Closest papers by meaning

• Scenario Dreamer: Vectorized Latent Diffusion for Generating Driving Simulation Environments (CVPR 2025 · similarity 0.87)
• Trajeglish: Traffic Modeling as Next-Token Prediction (ICLR 2024 · similarity 0.86)
• ChatScene: Knowledge-Enabled Safety-Critical Scenario Generation for Autonomous Vehicles (CVPR 2024 · similarity 0.86)
• SMART: Scalable Multi-agent Real-time Motion Generation via Next-token Prediction (NeurIPS 2024 · similarity 0.86)
• Flow Matching-Based Autonomous Driving Planning with Advanced Interactive Behavior Modeling (NeurIPS 2025 · similarity 0.86)
• DiffScene: Diffusion-Based Safety-Critical Scenario Generation for Autonomous Vehicles (AAAI 2025 · similarity 0.86)