Think2Drive: Efficient Reinforcement Learning by Thinking in Latent World Model for Quasi-Realistic Autonomous Driving

ECCV 24 2024 59.5 method, application

TLDR

Think2Drive uses a latent world model for efficient model-based RL to achieve expert-level autonomous driving in CARLA v2 within 3 days.

Reasoning

The paper introduces a novel model-based RL method with a world model for autonomous driving, achieving strong results in the challenging CARLA v2 simulator with high training efficiency. However, it lacks real-world validation and provides limited architectural details of the world model.

Read-first score

Read-first score 59.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 40.

Methodology quality 25%

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

Recency 8%

75.1

Uses a gentle age decay so recent papers surface without erasing older foundations. 2024

Topical relevance 42%

57.1

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

Reproducibility 25%

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

Field roles

Methodology anchor

Rank sensitivity

Stability: volatile; rank range: 120.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

world simulator

generative world model

interactive world model

video world model

Deep Analysis

Innovations

First model-based reinforcement learning method for autonomous driving (Think2Drive) using a world model as a neural simulator to train a planner.
CornerCase-Repository, a benchmark for evaluating driving models by scenarios.
A new balanced metric for driving performance combining route completion, infraction number, and scenario density.

Methodology

Think2Drive employs a model-based RL approach where a world model learns the environment transitions and then acts as a neural simulator to train a planner. The method leverages low-dimensional state space and parallel tensor computing in the world model to boost training efficiency. Training is conducted on the CARLA v2 simulator.

Key Results

Think2Drive achieves expert-level proficiency in CARLA v2 within 3 days of training on a single A6000 GPU, and is the first reported method to achieve 100% route completion on CARLA v2.

Resources

Paper