BridgeDrive: Diffusion Bridge Policy for Closed-Loop Trajectory Planning in Autonomous Driving

TL;DR

BridgeDrive presents a diffusion-bridge policy for autonomous driving that uses anchor-guided diffusion while preserving the symmetry between forward diffusion and denoising. By defining anchors as geometric-path waypoints and formulating a two-step generative process with an anchor-conditioned diffusion bridge, it trains a denoiser in a simulation-free setting and employs an anchor classifier at inference. The method achieves state-of-the-art performance on the Bench2Drive closed-loop benchmark, significantly improving success rate and driving score over prior diffusion-based planners, and operates with real-time capable ODE solvers. Limitations include dependence on LiDAR inputs and challenges in out-of-distribution scenarios, suggesting future directions such as VLA integration, distillation to one-step planning, and reinforcement-learning enhancements.

Abstract

Diffusion-based planners have shown great promise for autonomous driving due to their ability to capture multi-modal driving behaviors. However, guiding these models effectively in reactive, closed-loop environments remains a significant challenge. Simple conditioning often fails to provide sufficient guidance in complex and dynamic driving scenarios. Recent work attempts to use typical expert driving behaviors (i.e., anchors) to guide diffusion models but relies on a truncated schedule, which introduces theoretical inconsistencies and can compromise performance. To address this, we introduce BridgeDrive, a novel anchor-guided diffusion bridge policy for closed-loop trajectory planning. Our approach provides a principled diffusion framework that effectively translates anchors into fine-grained trajectory plans, appropriately responding to varying traffic conditions. Our planner is compatible with efficient ODE solvers, a critical factor for real-time autonomous driving deployment. We achieve state-of-the-art performance on the Bench2Drive benchmark, improving the success rate by 7.72% over prior arts.

Figures (7)

• Figure 1: Visualization of the denoising process of BridgeDrive ($t=T\to0$ from left to right), with the leftmost figure being anchor $x_T$ and the rightmost being the planned trajectory $x_0$. In each figure, the blue solid line depicts the denoised trajectory of the selected anchor at a specific timestep $t$, the red solid line depicts an example of the denoised trajectory of an un-selected anchor, and the rest scattered dots of other colors depict the denoised trajectories of other anchors at the timestep $t$. The red trajectory illustrates a failed case when a catastrophically wrong anchor is selected.
• Figure 2: Diagram for the planning procedure of BridgeDrive in \ref{['alg:sample']}. The model architecture of the neural network denoiser $x_\theta(x_t,t,x_T,z)$ is detailed in the light blue box.
• Figure 3: A consecutive four frames of a sample Bench2Drive scene, overtaking maneuver performed by BridgeDrive$^{\text{temp}}$. The ego car exhibited deficiencies in overtaking maneuver coordination and speed control, which directly led to a collision with the white vehicle. For video demonstration, please refer to supplementary materials.
• Figure 4: On the same scene as in \ref{['fig:overtaking_temp']}, overtaking maneuver performed by BridgeDrive$^{\text{geo}}$. The ego vehicle adapts its planning to overtake a sequence of parked cars. For video demonstration, please refer to supplementary materials.
• Figure 5: Full Diffusion model in a consecutive four frames of a sample Bench2Drive scene, failing to adhere to the target time window for lane-changing maneuvers, which consequently led to a collision with the road barrier. For video demonstration, please refer to supplementary materials.