autonomousvision/carla_garage

What it solves

It provides a comprehensive starter kit for the CARLA leaderboard 2.0, enabling researchers to develop, train, and evaluate end-to-end autonomous driving models. It addresses the difficulty of getting started with the CARLA simulator by providing open-source components including datasets, expert driver code, and training pipelines.

How it works

The project implements a framework for training autonomous agents using a multi-stage process. It utilizes the PDM-Lite expert driver to generate training data and provides pre-trained weights for TransFuser++, a high-performing open-source model. The system supports local evaluation, benchmarking against standard routes (such as Bench2Drive and Longest6 v2), and submission to the official CARLA leaderboard via Docker containers.

Who it’s for

It is designed for researchers and developers working on embodied intelligence and autonomous driving who want to benchmark their models against the CARLA leaderboard 2.0 standards.

Highlights

• Complete Starter Kit: Includes everything from data generation and training code to evaluation scripts.
• Pre-trained Models: Provides weights for TransFuser++, one of the best open-source models for this task.
• Multi-Benchmark Support: Integrated support for Bench2Drive, CARLA leaderboard 2.0 validation/test routes, and Longest6 v2.
• Open-Source Dataset: Includes a downloadable dataset used to train the released models.
• Guidance on Benchmarking: Provides a detailed guide on common methodological mistakes in autonomous driving benchmarking.