Diffusion-based driving planners capture diverse behaviors but may produce unsafe trajectories under distribution shift. To address this, we introduce BridgeGuard, a safety-constrained diffusion planning method. It progressively strengthens a constraint term during denoising, steering intermediate trajectories toward a scene-dependent safety domain. Corrections operate in a low-dimensional curve space, preserving geometric coherence.
We develop DistanceFieldNet, which takes bird's-eye-view features and predicts a time-dependent distance field. By sampling queries beyond expert trajectories and supervising both value and spatial gradient, the network learns the layout of safe and unsafe regions. The learned field supplies the constraint term during safety injection while the pretrained perception backbone and planner remain frozen.
We also derive sufficient conditions for terminal safety on an idealized continuous-time bridge. Bench2Drive experiments demonstrate cross-model generalization: for BridgeDrive, driving score and success rate rise from 87.99%/74.99% to 90.88%/76.36%; for $\text{DiffusionDrive}^{\text{geo}}$, they improve from 80.79%/58.18% to 90.46%/74.09%.
Review: BridgeGuard balances safety and diversity by embedding scene-aware constraints into the diffusion denoising process, achieving notable gains in realistic autonomous driving benchmarks.