← 研踪 PaperTrailarXiv:2609.39151 · v2
Linear Recurrent Memory Suffices to Distil a World-Model Policy for Robot Air Hockey
首次提交 2026-09-30 15:17
版本更新 2026-11-15 19:44
雷达观测 2026-11-17 19:15
摘要
Does memory-dependent control need nonlinear recurrent dynamics? We study simulated air-hockey defence under temporary loss of puck tracking. A DreamerV3 teacher outperforms a memoryless policy under tracking loss, and erasing its recurrent state at the moment tracking fails cuts its save rate sharply, so the task requires memory. We distil this teacher into compact recurrent policies with a 64-dimensional state, combining a diagonal linear recurrence with an optional rank- nonlinear innovation while the observation encoder and action head stay nonlinear. Across five matched seeds, the purely linear recurrent model () matches both the GRU baseline and the teacher throughout the tested range of tracking loss. Increasing nonlinear innovation rank provides no measured benefit. This result comes from a fresh test split that was opened only after all models and analyses were frozen. Under 5\,mm observation noise, without retraining, the linear student keeps a higher save rate than the GRU baseline, while using fewer recurrent parameters and less computation. These results suggest that, for this memory-dependent control task, nonlinear representation learning around a simple linear memory can be sufficient, and that nonlinear recurrent dynamics are not necessarily required.
作者备注:10 pages, 4 figures
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v2 · 2026-11-15 19:44 · 最新观测版本
Linear Recurrent Memory Suffices to Distil a World-Model Policy for Robot Air Hockey
Does memory-dependent control need nonlinear recurrent dynamics? We study simulated air-hockey defence under temporary loss of puck tracking. A DreamerV3 teacher outperforms a memoryless policy under tracking loss, and erasing its recurrent state at the moment tracking fails cuts its save rate sharply, so the task requires memory. We distil this teacher into compact recurrent policies with a 64-dimensional state, combining a diagonal linear recurrence with an optional rank- nonlinear innovation while the observation encoder and action head stay nonlinear. Across five matched seeds, the purely linear recurrent model () matches both the GRU baseline and the teacher throughout the tested range of tracking loss. Increasing nonlinear innovation rank provides no measured benefit. This result comes from a fresh test split that was opened only after all models and analyses were frozen. Under 5\,mm observation noise, without retraining, the linear student keeps a higher save rate than the GRU baseline, while using fewer recurrent parameters and less computation. These results suggest that, for this memory-dependent control task, nonlinear representation learning around a simple linear memory can be sufficient, and that nonlinear recurrent dynamics are not necessarily required.
查看此版本解读 →查看此版本原文 ↗