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

Authors: Aizierjiang Aiersilan

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments show that motion planners trained with our data significantly outperform those trained solely on realworld data, showing the usefulness of our synthetic data and the effectiveness of our data generation method.
Researcher Affiliation Academia Aizierjiang Aiersilan University of Macau EMAIL
Pseudocode Yes Algorithm 1: Filter for Scenario Descriptions Algorithm 2: Validator for Generated Configuration
Open Source Code Yes Code https://ezharjan.github.io/Auto Scene Gen
Open Datasets Yes While existing works on vehicle or pedestrian trajectory prediction, such as Traffic Predict (Ma et al. 2019), Pishgu (Alinezhad Noghre et al. 2023), Tra PHic (Chandra et al. 2019), and MSRL (Wu et al. 2023), rely on real-world datasets like Apollo Scapes (Huang et al. 2018), NGSIM (Kovvali, Alexiadis, and Zhang PE 2007), UCY (Lerner, Chrysanthou, and Lischinski 2007), and ETH (Pellegrini et al. 2009)...
Dataset Splits No The paper states: "We generated 17,919 examples; the official training set of A.S. contains 94 examples." and "By directly replacing only the train set of the target dataset, we initiated the evaluation process..." However, it does not provide specific train/test/validation splits for its own generated dataset or explicit percentages for the datasets it augments or replaces.
Hardware Specification Yes Data collection was performed on a machine with an NVIDIA GeForce RTX 3090 GPU, 12th Gen Intel(R) Core(TM) i7-12700, using CARLA version 0.9.13 with the Town03 map (300 300).
Software Dependencies Yes Data collection was performed on a machine with an NVIDIA GeForce RTX 3090 GPU, 12th Gen Intel(R) Core(TM) i7-12700, using CARLA version 0.9.13 with the Town03 map (300 300). ... we used Auto Scene Gen with GPT-4 (Achiam et al. 2023) as the selected LLM to generate 125 traffic scenarios...
Experiment Setup No The paper mentions: "We replicated Traffic Predict s results (Ma et al. 2019) with their Apollo Scapes dataset..." and "Without modifying the original trajectory prediction network, our dataset achieved superior results...". This implies using the experimental setup from cited works, but specific hyperparameters, learning rates, batch sizes, or optimizer settings are not explicitly detailed within this paper's main text.