⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

Learning to drive from a world on rails

arXiv 2021 51.7 method, application

TLDR

Proposes a model-based driving policy with a world-on-rails assumption, outperforming IL and RL on CARLA and ProcGen benchmarks.

Reasoning

The paper introduces a novel 'world on rails' assumption that simplifies learning by factorizing dynamics into a non-reactive world model and a compact forward model, allowing effective training from pre-recorded logs. Strengths include strong empirical results on standard benchmarks and sample efficiency, but the assumption may limit applicability to environments where agent actions significantly affect the world.

Read-first score

Read-first score 51.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 38.

Methodology quality 25%
70

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

Topical relevance 42%
54.3

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

Recency 8%
49

Uses a gentle age decay so recent papers surface without erasing older foundations. 2021

Reproducibility 25%
30

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

Methodology anchor

Rank sensitivity

Stability: volatile; rank range: 174.

Keyword Scores

world model
9
model-based reinforcement learning world model
9
world dynamics prediction
8
world simulator
6
interactive world model
3
generative world model
2
video world model
1

Deep Analysis

Innovations

  • World-on-rails assumption: agent's actions do not influence the environment, simplifying learning from pre-recorded logs.
  • Factorizing dynamics into a non-reactive world model and a low-dimensional forward model of the ego-vehicle.
  • Using tabular dynamic-programming on Bellman equations to compute action-values for training trajectories, then supervising a vision-based policy.

Methodology

A model-based approach learns a vision-based driving policy from pre-recorded logs under the assumption that the world is on rails, meaning the environment is non-reactive. The dynamics are factorized into a static world model and a compact ego-vehicle forward model, and action-values are computed via tabular dynamic programming on the Bellman equations to supervise the policy.

Key Results

The policy outperforms imitation learning, model-based, and model-free reinforcement learning on CARLA NoCrash, and is an order of magnitude more sample-efficient than state-of-the-art model-free RL on ProcGen navigational tasks.

Limitations

  • Relies on the world-on-rails assumption, meaning the environment is treated as non-reactive during training; applicability may be limited when logs do not capture full environmental interactions.

Resources

Paper

Tags