Mitigating Covariate Shift in Imitation Learning for Autonomous Vehicles Using Latent Space Generative World Models

arXiv 24.9 2024 39.3 method, application

TLDR

Proposes latent space generative world models to mitigate covariate shift in autonomous driving imitation learning, with simulator results.

Reasoning

The paper directly addresses covariate shift using world models and presents novel perception encoder, with strong simulator results. However, it lacks real-world validation and the world model is not used for interactive or video tasks, limiting scope.

Read-first score

Read-first score 39.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 36.

Recency 6%

75.1

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

Methodology quality 18%

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

Topical relevance 29%

51.4

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 18%

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

Citation impact 18%

10.1

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=1

Citation velocity 12%

5.9

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.33

Field roles

Candidate

Rank sensitivity

Stability: volatile; rank range: 202.

Keyword Scores

world model

generative world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

video world model

Deep Analysis

Innovations

Use of latent space generative world models to address covariate shift in imitation learning for autonomous driving
Novel transformer-based perception encoder employing multi-view cross-attention and a learned scene query

Methodology

The paper proposes using a latent space generative world model, a neural network that predicts the agent's next state given past states and actions, to mitigate covariate shift. During end-to-end training, the driving policy learns to recover from errors by aligning with states observed in human demonstrations. The approach also introduces a transformer-based perception encoder with multi-view cross-attention and a learned scene query. Training and evaluation are conducted in the CARLA simulator and NVIDIA's DRIVE Sim.

Key Results

The method demonstrates significant improvements upon prior state of the art in closed-loop testing in the CARLA simulator, and shows the ability to handle perturbations in both CARLA and NVIDIA's DRIVE Sim.

Resources

Paper