arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2604.14484cs.ROcs.AImath.OC
收藏

行为克隆中增益依赖误差动态的非渐近理论

Behavior Cloning Under PD Control: A Finite-Horizon Theory of Gain-Dependent Error Amplification

PIT IN Corp.(PIT IN公司)

Junghoon Seo

AI总结:

本文研究了行为克隆中控制器增益对有限时间任务失败的影响,通过分析误差传播机制,推导出代理矩阵X_∞(K)的表达式,并提出基于结构假设的界限,揭示了增益依赖误差衰减的非渐近有限时间扩展。

摘要:

行为克隆(BC)策略在位置控制机器人上继承了底层PD控制器的闭环响应,但控制器增益对BC失败的非渐近有限时间后果仍不明确。本文显示,独立的子高斯动作误差通过增益依赖的闭环动态传播,产生子高斯位置误差,其代理矩阵X_∞(K)控制失败尾部。任务失败概率T时间因子分解为增益依赖的放大指数Γ_T(K)和验证损失加一般化松弛,因此训练损失单独不能预测闭环性能。在保持上界结构假设下,代理矩阵允许标量界X_∞(K)≤Ψ(K)X̄,其中Ψ(K)分解为标签难度、注入强度和收缩。这对四个典型区域进行排序,合规过阻尼(CO)最紧,刚性欠阻尼(SU)最松,刚性过阻尼与合规欠阻尼的顺序系统依赖。对于典型标量二阶PD系统,闭环连续时间稳态方差X_∞^c(α,β)=σ²α/(2β)在稳定正交象限内严格单调递增,涵盖欠阻尼和过阻尼区域,且精确零阶保持(ZOH)离散化继承此单调性。分析给出了Bronars等人的增益依赖误差衰减解释的非渐近有限时间扩展。

英文摘要:

Behavior cloning (BC) on position-controlled robots is shaped by the PD loop that executes policy actions. We give a finite-horizon, nonasymptotic analysis of how controller gains affect BC failure. Independent sub-Gaussian action errors propagate through gain-dependent closed-loop dynamics into sub-Gaussian position errors. The resulting failure tail is controlled by controller amplification multiplied by validation loss and generalization slack, so validation loss alone can mis-rank gains. Under shape-preserving upper-bound assumptions, the analysis separates label difficulty, injection strength, and contraction, ranking compliant-overdamped gains as tightest and stiff-underdamped gains as loosest, with the mixed regimes system-dependent. In the canonical scalar second-order PD system, stationary position-error variance increases with stiffness and decreases with damping over the stable range, and exact zero-order-hold discretization inherits the ordering to leading order. This extends the error-attenuation explanation of bronars et al. (2026) to finite-horizon failure bounds.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑