跳转至

文章背景与核心概要

部署在嵌入式系统上的学习型自动驾驶策略通常会经历各种模型压缩技术（如剪枝、知识蒸馏和量化），以满足内存和功耗的限制。然而，传统的评估通常依赖于宏观的数值评分，这往往无法捕捉到在与其他道路使用者进行闭环交互过程中的现实世界安全风险。

本文引入了一种分阶段的闭环评估方法，以追踪模型压缩对驾驶能力的影响。作者将驾驶任务建模为部分可观测马尔可夫决策过程（POMDP），并在 Gym-Duckietown 中通过近端策略优化（PPO）进行训练，随后在各个压缩阶段对智能体通过五个驾驶课程进行了评估：结构化剪枝被认为是驾驶能力首次丢失的初始失效点；知识蒸馏部分恢复并改进了被剪枝的智能体，尽管受限于其排练数据的局限性；整型量化在应用于改进后的智能体时，会导致特定需要停车和恢复机动的课程出现性能退化。相反，将量化直接应用于未剪枝的智能体则成功保留了全部五个驾驶课程的能力。

A Closed-Loop Evaluation of Capability Loss and Recovery in Compressed Driving Policies

arXiv: 2609.00718 [cs.AI]
Submitted: September 1, 2026
Authors: Ahmad Alfan Alfian Irfan, Nur Ahmad Khatim, Mansur Arief

arXiv: 2609.00718 [cs.AI]
Submitted: September 1, 2026
Authors: Ahmad Alfan Alfian Irfan, Nur Ahmad Khatim, Mansur Arief

📌 Summary

Learned autonomous driving policies deployed on embedded systems often undergo compression techniques—such as pruning, knowledge distillation, and quantization—to meet memory and power constraints. However, traditional evaluations rely on aggregate numerical scores that often fail to capture real-world safety risks during closed-loop interactions with other road users.

This paper introduces a stage-wise closed-loop evaluation approach to track how compression impacts driving capability. Formulating the driving task as a Partially Observable Markov Decision Process (POMDP) trained via Proximal Policy Optimization (PPO) in Gym-Duckietown, the authors evaluate actors across five driving curricula through each compression stage:

  • Structured Pruning: Identified as the initial failure point where driving capabilities are first lost.
  • Knowledge Distillation: Partially recovers and improves the pruned actor, though constrained by the limitations of its rehearsal data.
  • Integer Quantization: Causes regression on specific curricula that require stopping and resuming maneuvers when applied to the improved actor. Conversely, applying quantization directly to the unpruned actor successfully preserves all five driving curricula.

📌 Summary

Learned autonomous driving policies deployed on embedded systems often undergo compression techniques—such as pruning, knowledge distillation, and quantization—to meet memory and power constraints. However, traditional evaluations rely on aggregate numerical scores that often fail to capture real-world safety risks during closed-loop interactions with other road users.

This paper introduces a stage-wise closed-loop evaluation approach to track how compression impacts driving capability. Formulating the driving task as a Partially Observable Markov Decision Process (POMDP) trained via Proximal Policy Optimization (PPO) in Gym-Duckietown, the authors evaluate actors across five driving curricula through each compression stage:

  • Structured Pruning: Identified as the initial failure point where driving capabilities are first lost.
  • Knowledge Distillation: Partially recovers and improves the pruned actor, though constrained by the limitations of its rehearsal data.
  • Integer Quantization: Causes regression on specific curricula that require stopping and resuming maneuvers when applied to the improved actor. Conversely, applying quantization directly to the unpruned actor successfully preserves all five driving curricula.