Automatic Curricula via Expert Demonstrations (ACED)

ACED constructs a curriculum by sampling states from expert demonstration trajectories as initializations for each training episode, where the samples initially come from near the end of the demonstration trajectories and gradually move forward as the agent improves its performance.

ACED 方法通过从专家演示轨迹中采样状态，并将其作为每个训练回合的初始化来构建课程式学习流程。其中，初始采样的状态最初源自演示轨迹的末端区域，随着智能体性能不断提升，采样的起始位置会逐步向轨迹前端移动。

搜集汇总

背景与挑战

背景概述

ACED数据集利用专家演示轨迹构建自动课程，通过从轨迹末端向前逐步采样的方式初始化训练episode，从而根据智能体表现动态调整训练难度。