PrismAD: Decoupled Planning via Semantic Mixture-of-Planners for End-to-End Autonomous Driving

2026-07-11

AI Summary

This work addresses the limitations of existing end-to-end autonomous driving planners, which couple heterogeneous scene information into a unified representation space, thereby hindering specialized modeling of critical factors such as agent interactions, road geometry, and driving intent. To overcome this, the authors propose PrismAD, a novel semantic-driven decoupled planning framework. PrismAD introduces a semantic mixture planner that decomposes inputs into three distinct semantic groups—interaction, geometry, and intent—each processed by dedicated expert networks with disjoint parameters. A semantic-aware router then adaptively fuses these outputs using sparse Top-K activation combined with a noisy gating mechanism. This design enhances routing robustness while reducing computational overhead, achieving competitive performance on both the nuScenes open-loop dataset and the NeuroNCAP closed-loop benchmark.

Abstract

This letter presents PrismAD, a decoupled end-to-end autonomous driving framework based on a Semantic Mixture-of-Planners. Existing planners usually aggregate heterogeneous scene tokens into a coupled representation space, forcing a single planning branch to jointly model agent interaction, road geometry, and driving intention. Such coupling may weaken factor-specific reasoning and obscure the contribution of different planning cues. To address this limitation, PrismAD partitions scene tokens into interaction, geometry, and intent groups, and assigns them to independent planning experts with the same architecture but separate parameters. Each expert learns a specialized motion-planning representation, while a semantics-aware router adaptively aggregates expert predictions with separate routing weights for motion prediction and ego planning. Sparse top-$K$ activation with noisy gating is further introduced to improve routing robustness and reduce unnecessary expert computation. Extensive experiments on the nuScenes open-loop dataset and NeuroNCAP closed-loop benchmark demonstrate that PrismAD exhibits competitive performance. Our code will be released soon.

Problem

• autonomous driving
• motion planning
• scene representation
• planning decoupling
• semantic reasoning

Innovation

• Semantic Mixture-of-Planners
• Decoupled Planning
• End-to-End Autonomous Driving
• Sparse Routing
• Expert Specialization