Diffusion Forcing Planner: History-Annealed Planning with Time-Dependent Guidance for Autonomous Driving

Researchers propose Diffusion Forcing Planner (DFP), a new diffusion-based motion planning framework for autonomous driving that addresses temporal inconsistency in learning-based planners. By decomposing trajectories into history, current, and future segments with independent noise levels and applying annealed guidance, DFP produces more stable and controllable driving plans while avoiding the tendency to simply copy historical patterns.

The autonomous driving industry faces a critical technical challenge: learning-based motion planners generate temporally inconsistent trajectories where small frame-to-frame perturbations accumulate into unsafe, uncomfortable driving behaviors. While previous attempts to stabilize output by conditioning on historical data succeeded in reducing jitter, they created a new problem—models learned to replicate past patterns rather than adapt dynamically to environmental changes.

DFP represents an architectural innovation addressing this core tension. Rather than treating history as a static conditioning input, the framework applies heterogeneous diffusion processes across trajectory segments. By assigning independent noise levels to history, present, and future components and jointly denoising them, the model learns to maintain temporal continuity while preserving adaptive capacity. The classifier-free guidance mechanism enables annealed control over future sampling, allowing the planner to respect historical momentum without becoming locked into repetitive behavior patterns.

The competitive performance on nuPlan's closed-loop evaluation benchmarks suggests this approach successfully resolves the stability-versus-adaptability tradeoff that has limited prior motion planning methods. This matters for both autonomous vehicle developers facing real-world deployment challenges and the broader diffusion model community exploring applications beyond generative tasks.

The technical contribution—decomposed noise scheduling in diffusion processes—may find applications across sequential decision-making domains beyond driving. Future work likely explores scaling to longer planning horizons and integration with perception uncertainty quantification.

• →DFP solves temporal inconsistency in autonomous driving planners by applying independent noise levels to history, current, and future trajectory segments.
• →The framework uses annealed classifier-free guidance to enable controllable, stable motion planning without sacrificing environmental adaptation.
• →Closed-loop evaluation on nuPlan demonstrates competitive performance while maintaining trajectory smoothness in complex driving scenarios.
• →Heterogeneous diffusion processes represent a transferable architectural pattern potentially applicable to other sequential decision-making tasks.
• →The approach balances historical conditioning benefits against the risk of pattern-copying that plagued previous history-augmented planners.