Solving Motion Planning Tasks with a Scalable Generative Model
TLDR
A scalable generative model for autonomous driving that learns scene dynamics, enabling simulation, scenario generation, and planning, achieving SOTA on real-world benchmarks.
Reasoning
Strengths include novel autoregressive modes for efficiency, strong empirical results on two real-world datasets, and open-source code. Weaknesses are limited to driving domain and lack of discussion on generalization or long-horizon simulation limitations.
Read-first score
Read-first score 74.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 54.
Field roles
Rank sensitivity
Stability: volatile; rank range: 75.
Keyword Scores
Deep Analysis
Innovations
- Scalable generative model that learns dynamics of driving scenes for motion planning
- Dual-mode autoregressive design (full-autoregressive and partial-autoregressive) improving inference and training speed without sacrificing generative capability
- Ability to simulate diverse futures and generate driving scenarios conditioned on various prompts
- Unified model serving as online reactive environment for reinforcement learning, evaluator for planning policies, and high-fidelity simulator
Methodology
The paper proposes a generative model that learns the dynamics of driving scenes. It operates in both full-autoregressive and partial-autoregressive modes to enhance training and inference efficiency. The model is evaluated on the Waymo motion dataset and nuPlan dataset for simulation realism, scene generation, and planning benchmarks.
Key Results
The model achieves state-of-the-art performance on simulation realism and scene generation benchmarks, and the planner outperforms prior arts on planning benchmarks.