ASSCG: Just-Right Gating over Chattering for Fast-Slow LLM Planning in Autonomous Driving

2026-06-24

AI Summary

This work addresses the inefficiency of existing hierarchical planning frameworks, where handcrafted triggering rules often lead to either excessive or mistimed invocations of the slow planner, degrading both performance and computational efficiency. To overcome this limitation, the authors formulate slow-planner invocation as a resource-aware sequential decision-making problem and propose an Adaptive Slow System Control Gate (ASSCG) that dynamically determines, at the frame level, whether to query, cache, or discard guidance signals from the slow system. This approach explicitly incorporates computational resource constraints into the invocation policy and leverages an RWKV backbone for long-horizon gating decisions. The model is trained via supervised fine-tuning followed by a computation-aware GRPO-style reinforcement fine-tuning strategy. Experiments demonstrate that the method achieves a closed-loop score of 67.28 on nuPlan Hard20 with 60% lower inference latency, and attains a PDMS of 91.4 on NAVSIM with a 25% improvement in average speed.

Abstract

Large language models (LLMs) can improve autonomous driving planning but are costly to query online, and existing fast-slow planners often rely on hand-designed triggering rules that either over-call the slow system or call it at the wrong times. We formulate slow-system invocation as a resource-aware sequential decision problem and propose the Adaptive Slow-System Control Gate (ASSCG), which makes frame-level Query/Cache/Drop decisions to refresh, reuse, or suppress slow guidance. ASSCG uses an RWKV backbone for efficient long-horizon gating and is trained with supervised fine-tuning followed by GRPO-style compute-aware reinforcement fine-tuning. We apply ASSCG to two different fast-slow architectures: (i) AsyncDriver on nuPlan Hard20 closed-loop evaluation, where ASSCG improves score to 67.28 (+2.28) while reducing average end-to-end inference latency by 60%; and (ii) a RecogDrive-based dual system that we build by replacing its original VLM-2B module with a lightweight ViT-based fast planner and adding an LLM slow planner, evaluated on NAVSIM, where ASSCG achieves 91.4 PDMS (+0.6) and increases average speed by 25%. The project page, including video visualizations and additional results, is available at https://williamxuanyu.github.io/asscg/.

Problem

• autonomous driving
• large language models
• fast-slow planning
• slow-system invocation
• resource-aware decision

Innovation

• Adaptive Gating
• Fast-Slow Planning
• Resource-Aware Control
• LLM-based Autonomous Driving
• RWKV