Diffusion-based driving planners are attractive because they can capture a wide range of driving behaviors. However, under distribution shift—when the operating environment differs from training data—they may produce unsafe trajectories. That gap between flexibility and safety is the problem BridgeGuard targets.
BridgeGuard is a safety-constrained diffusion planning method. It works by progressively strengthening safety constraints during the planning process, rather than applying a fixed safety filter. This allows the planner to retain its behavioral diversity while steering generation away from dangerous outputs.
The approach is described in a new arXiv preprint. As a single-source digest, the details are limited to the abstract, but the core idea is clear: safety should be built into the diffusion process itself, not bolted on afterward.