Dream to Drive: Model-Based Vehicle Control Using Analytic World Models
TLDR
Proposes Analytic World Models using differentiable simulators for predictive, prescriptive, and counterfactual capabilities in autonomous driving.
Reasoning
The paper introduces a novel approach to world modeling by leveraging differentiable simulators for end-to-end learning of state predictors, odometry, planners, and inverse states. Its strengths lie in the innovative task setups and broad applicability in autonomous driving, but it lacks real-world validation, relying solely on simulated environments.
Read-first score
Read-first score 44.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 33.
Field roles
Rank sensitivity
Stability: volatile; rank range: 405.
Keyword Scores
Deep Analysis
Innovations
- Using differentiable simulators for world modeling (Analytic World Models) rather than just policy gradients
- Three novel task setups: learning relative odometry, optimal planners, and optimal inverse states via end-to-end computation graph combining differentiable dynamics with a state predictor
- End-to-end learning of state predictors through differentiable simulation, enabling predictive, prescriptive, and counterfactual capabilities
Methodology
The methodology combines differentiable dynamics within an end-to-end computation graph with a state predictor, not a policy. Three task setups are designed: relative odometry, optimal planners, and optimal inverse states. The models are trained via backpropagation through the differentiable simulator, enabling efficient learning of Analytic World Models (AWMs).
Key Results
The paper demonstrates that differentiable simulation enables efficient, end-to-end learning of Analytic World Models, which have broad applicability in autonomous driving scenarios and can augment decision-making beyond reactive control.