핵심 개념
The proposed Versatile Behavior Diffusion (VBD) model leverages diffusion-based generative modeling to efficiently generate realistic and controllable traffic scenarios by integrating joint multi-agent diffusion policy and marginal multi-modal behavior prediction.
초록
The paper introduces the Versatile Behavior Diffusion (VBD) model, which utilizes diffusion-based generative modeling to generate realistic and controllable traffic scenarios. The key components of VBD are:
- Scene Encoder: A query-centric Transformer-based encoder that encodes the states of agents and map polylines in their local coordinates, preserving relative information.
- Denoiser: A Transformer-based module that generates scene-level joint behaviors of agents from noise, enabling one-step generation and iterative refinement.
- Behavior Predictor: A Transformer-based module that forecasts multi-modal trajectories for individual agents as behavior priors, which can be used to guide the denoising process.
VBD demonstrates state-of-the-art performance on the Waymo Sim Agents benchmark for multi-agent interaction modeling. The model's versatility is showcased through various sampling strategies:
- One-step generation using the denoiser
- Multi-step sampling for improved diversity
- Guided sampling with user-specified constraints or targets, using cost functions or game-theoretic objectives
- Fusing behavior priors from the predictor to generate diverse and scene-consistent scenarios
The paper also provides insights on the connection between diffusion-based generative modeling and the classic imitation learning formulation of traffic scenario generation.
통계
The paper does not provide specific numerical data or statistics to support the key logics. The focus is on the conceptual framework and model architecture.
인용구
"Generating realistic and controllable agent behaviors in traffic simulation is crucial for the development of autonomous vehicles."
"We draw a conceptual connection between IL and diffusion-based generative modeling and introduce a novel framework Versatile Behavior Diffusion (VBD) to simulate interactive scenarios with multiple traffic participants."
"VBD not only generates scene-consistent multi-agent interactions but also enables scenario editing through multi-step guidance and refinement."