Top of the CARLA Leaderboard, With Commentary

What it does InterFuser is an end-to-end autonomous driving system for the CARLA simulator that ingests multi-view cameras, LiDAR, and bird’s-eye-view maps through a transformer backbone. It produces intermediate interpretable features—semantic representations that constrain the vehicle’s actions to remain within predicted safe sets rather than emitting raw control signals from an opaque black box. The authors report that it held state-of-the-art status on the CARLA AD Leaderboard as of June 2022.

The interesting bit Most sensor-fusion driving stacks treat perception as an inscrutable black box; InterFuser forces the network to generate human-readable scene semantics that double as hard safety guardrails. The implementation extends Transfuser and the 2020 CARLA challenge codebase, so the contribution is architectural rather than a ground-up reinvention of the surrounding scaffolding.

Key highlights

• Fuses multi-view RGB, segmentation, depth, LiDAR, and bird’s-eye-view inputs through a transformer
• Generates intermediate interpretable features to constrain steering and throttle within safe action sets
• Includes data generation, training, and evaluation scripts tied to CARLA 0.9.10.1
• Provides pre-trained weights, though the README notes they were trained on a subset of towns and weathers
• Achieved top rank on the CARLA AD Leaderboard (June 2022) per the authors

Caveats

• Pre-trained weights are trained on only part of the full dataset (several towns and weathers), so generalization is unclear
• Heavily coupled to the CARLA 0.9.10.1 ecosystem and its leaderboard infrastructure

Verdict Worth exploring if you research interpretable end-to-end driving or need a strong CARLA baseline with built-in safety semantics. Give it a pass if you need real-world deployment code or a simulator-agnostic framework.