Does End-to-End Autonomous Driving Really Need Perception Tasks?

ICLR 25 2025 40 method, application

TLDR

SSR uses sparse scene tokens to eliminate supervised perception in end-to-end autonomous driving, improving speed and safety.

Reasoning

The paper presents a novel framework that removes the need for expensive annotations by using navigation-guided sparse tokens, achieving significant improvements in efficiency and performance. However, it does not address world models or dynamics prediction, limiting its scope to direct driving tasks.

Read-first score

Read-first score 40, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 0.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=True; dataset=False; markers=code,github

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=benchmark

Topical relevance 42%

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Field roles

FrontierReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 299.

Keyword Scores

world model

world simulator

generative world model

interactive world model

video world model

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Eliminates need for human-designed supervised perception sub-tasks in end-to-end autonomous driving
Introduces Sparse Scene Representation (SSR) using only 16 navigation-guided tokens to extract crucial scene information
Temporal enhancement module that aligns predicted future scenes with actual future scenes via self-supervision

Methodology

SSR uses 16 navigation-guided tokens as a sparse scene representation, eliminating supervised perception tasks. A temporal enhancement module is introduced to align predicted future scenes with actual future scenes through self-supervision, allowing computational resources to focus on navigation-relevant elements.

Key Results

SSR achieves a 27.2% relative reduction in L2 error and a 51.6% decrease in collision rate compared to UniAD on nuScenes, with 10.9x faster inference and 13x faster training. It also outperforms VAD-Base by 48.6 points on driving score in CARLA's Town05 Long benchmark.

Resources

Paper Code