Learning Multiple Probabilistic Decisions from Latent World Model in Autonomous Driving

arXiv 24.9 2024 72.2 method, application

TLDR

LatentDriver uses a latent world model with mixture distributions to handle uncertainty and self-delusion in autonomous driving, outperforming SOTA on Waymax.

Reasoning

The paper addresses key challenges in autoregressive world models for decision-making by modeling multiple probabilistic hypotheses, which is a novel contribution. Strengths include clear problem formulation and strong empirical results on a closed-loop benchmark. Weaknesses are limited scope (only vectorized scenes, no video) and lack of ablation on mixture modeling components.

Read-first score

Read-first score 72.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 39.

Methodology quality 25%

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

Reproducibility 25%

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

Recency 8%

75.1

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

Topical relevance 42%

55.7

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

Methodology anchorReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 225.

Keyword Scores

world model

world dynamics prediction

generative world model

world simulator

interactive world model

model-based reinforcement learning world model

video world model

Deep Analysis

Innovations

Formulation of multiple probabilistic hypotheses to address insufficient uncertainty modeling in autoregressive world models
LatentDriver framework that models environment next states and ego vehicle actions as a mixture distribution
Mitigation of self-delusion problem by providing intermediate actions sampled from a distribution to the world model

Methodology

LatentDriver uses an autoregressive world model to capture vectorized scene understanding, then models the environment's next states and the ego vehicle's possible actions as a mixture distribution. A deterministic control signal is derived from this mixture, and the self-delusion issue is reduced by feeding sampled intermediate actions back into the world model.

Key Results

On the Waymax closed-loop benchmark, LatentDriver surpasses state-of-the-art reinforcement learning and imitation learning methods, achieving expert-level performance.

Limitations

Self-delusion problem is mitigated but not fully eliminated
Evaluation is limited to the Waymax benchmark; generalization to other driving scenarios is not demonstrated
Reliance on autoregressive world model may inherit its generalization limitations and computational costs

Resources

Paper Code