PLAN-S: Bridging Planning with Latent Style Dynamics for Autonomous Driving World Models

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

PLAN-S, a novel planner-facing bridge, enhances latent world models (LWMs) for autonomous driving by explicitly decoding a style-conditioned, four-channel semantic cost map from latent representations. This cost map, covering dynamic obstacles, off-road regions, static obstacles, and drivability, is conditioned on ego state and driving style via a dual AdaFiLM mechanism. PLAN-S integrates with regression planners through attention-level fusion and with anchor-score planners via reward-level fusion, keeping host backbones frozen. Validated on ResWorld with nuScenes and WoTE with NAVSIM, PLAN-S reduced L2 error to 0.55 m and the 3 s collision rate by 42% on nuScenes. On NAVSIM, the rule-cost variant achieved 89.4 Predictive Driver Model Score (PDMS), with the learned cost variant providing complementary gains on challenging scenes. The system adds only 0.25 million parameters and runs at 17.0 frames per second.

Key takeaway

For Machine Learning Engineers developing autonomous driving systems, PLAN-S offers a robust method to integrate explicit style-conditioned spatial costs into latent world models. You should consider implementing a similar planner-facing bridge to enhance trajectory safety and interpretability, especially for collision reduction. This approach allows for inspectable risk modeling and diverse style preferences, improving performance on challenging driving scenarios without significant inference overhead.

Key insights

PLAN-S enhances autonomous driving LWMs by explicitly modeling style-conditioned spatial costs for improved controllability and safety.

Principles

• Explicitly model risk, drivability, and style preferences.
• Organize latent representations as spatial costs.
• Ensure portability across planner families.

Method

PLAN-S decodes a four-channel semantic cost map from BEV latent features, conditioned by ego state and driving style via dual AdaFiLM. It integrates via attention-level fusion for regression planners or reward-level fusion for anchor-score planners.

In practice

• Use a 4-channel cost map for explicit risk/drivability.
• Apply dual AdaFiLM for style-conditioned modulation.
• Integrate cost maps before final trajectory selection.

Topics

• Autonomous Driving
• Latent World Models
• Semantic Cost Maps
• Driving Style Personalization
• Trajectory Planning
• Deep Learning Architectures

Best for: Computer Vision Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• PLAN-S: Bridging Planning with Latent Style Dynamics for Autonomous Driving World Models — PLAN-S improves autonomous driving LWMs by using style-conditioned semantic cost maps for explicit risk and style modulation in planning.
• CLEAR: Cognition and Latent Evaluation for Adaptive Routing in End-to-End Autonomous Driving — CLEAR achieves efficient, multi-modal autonomous driving by combining single-step generative planning with LLM-driven cognitive reasoning.
• Latent-WAM: Latent World Action Modeling for End-to-End Autonomous Driving — Latent-WAM improves autonomous driving planning via efficient, spatially-aware, and dynamics-informed latent world models.
• Xiaomi EV’s Next Self-Driving AI Explained — Autonomous driving models can achieve superior performance by shifting from linguistic reasoning to a latent spatial-temporal mathematical representation.
• Beyond Waypoints: A Trajectory-Centric Waypointing Paradigm for Vision-Language Navigation — Grounding waypoints in executable trajectories resolves planning-control inconsistencies and improves reachability in Vision-Language Navigation.
• Multi-Agent Reinforcement Learning for Safe Autonomous Driving Under Pedestrian Behavioral Uncertainty — Co-training SDCs with personality-driven MARL pedestrians significantly improves safety by exposing vehicles to realistic, uncertain human behaviors.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.