LfGP Data (Learning from Guided Play Expert Data and Trained Models)

从引导游戏中学习: 一种预定的分层方法，用于改善对抗性模仿学习源代码中的探索
对抗模仿学习 (AIL) 是一种从演示中学习的技术，有助于纠正监督学习 (行为克隆) 中出现的分布转移问题。在我们的论文中，我们通过许多实验和分析表明，在操纵环境中，AIL遭受欺骗性奖励，从而导致次优策略。在这项工作中，除了主要任务外，我们还通过对一组易于定义的辅助任务进行探索来解决此问题。

Learning from Guided Play: A Prescribed Hierarchical Approach to Enhancing Exploration in Adversarial Imitation Learning Source Code

Adversarial Imitation Learning (AIL) is a demonstration-based learning technique that helps mitigate distribution shift issues arising in supervised learning (Behavior Cloning). In this paper, we demonstrate through extensive experiments and analyses that AIL suffers from deceptive rewards in manipulation environments, leading to suboptimal policies. In this work, we address this problem by exploring a set of readily definable auxiliary tasks in addition to the primary task.

搜集汇总

背景与挑战

背景概述

该数据集名为LfGP Data，专注于对抗模仿学习领域，旨在通过引入辅助任务来解决欺骗性奖励导致的次优策略问题，从而改善探索过程。它由多伦多大学于2021年发布，相关资源包括源代码和论文。