From 2022.3 to present. UCLA Prof Bolei-Zhou’s group.
- Latest news: The work has been accept by
NeuralIPS 2023 track on Datasets and Benchmarks! Camera-ready paper will be released shortly.
ScenarioNet is an open-sourced platform for large-scale traffic scenario modeling and simulation. ScenarioNet defines a unified scenario description format containing HD maps and detailed object annotations. Scenarios recorded in this format can be replayed in the digital twins with multiple views, ranging from Bird-Eye-View layout to realistic 3D rendering. ScenarioNet provides tools to build and manage databases built from various data sources including real-world datasets like Waymo, nuScenes, Lyft L5, and nuPlan datasets and synthetic datasets like the procedural generated ones and safety-critical ones. We demonstrate several applications of ScenarioNet including large-scale scenario generation, AD testing, imitation learning, and reinforcement learning in both single-agent and multi-agent settings. The results imply scaling up the training data brings new research opportunities in machine learning and autonomous driving.
ScenarioNet consists of the data layer, system layer, and application layer. Various datasets are unified into an internal scenario description. The system layer then provides a set of tools to operate on data efficiently, such as filtering, merging, sanity-check, splitting and so on. Once the database is ready, it can be loaded into MetaDrive for large-scale simulation and supports applications.