CMU Play Fusion

从非结构化和未经整理的数据中学习已成为语言或视觉生成方法的主导范例。这种非结构化和无指导的行为数据（通常称为游戏）在机器人技术中也更容易收集，但由于其固有的多模式、噪声和次优性质，学习起来更加困难。在本文中，我们研究了从事后用语言标记的非结构化游戏数据中学习目标导向的技能策略的问题。具体来说，我们利用扩散模型的进步来学习多任务扩散模型，以从游戏数据中提取机器人技能。在状态和动作空间中使用条件去噪扩散过程，我们可以优雅地处理游戏数据的复杂性和多模态性，并生成多样化且有趣的机器人行为。为了使扩散模型对技能学习更有用，我们鼓励机器人代理通过在条件行为生成过程中引入离散瓶颈来获取技能词汇。在我们的实验中，我们展示了我们的方法在模拟和现实世界的各种环境中的有效性。

Learning from unstructured and uncurated data has become the dominant paradigm for language or vision generation methods. This unstructured and unguided behavioral data, often referred to as "games", is also easier to collect in robotics, but it is more difficult to learn from due to its inherent multimodal, noisy, and suboptimal nature. In this paper, we investigate the problem of learning goal-oriented skill policies from unstructured game data labeled with language post-hoc. Specifically, we leverage advances in diffusion models to learn a multi-task diffusion model for extracting robotic skills from game data. By employing a conditional denoising diffusion process over both state and action spaces, we can elegantly handle the complexity and multimodality of game data, and generate diverse and engaging robotic behaviors. To make diffusion models more useful for skill learning, we encourage robotic agents to acquire a skill vocabulary by introducing a discrete bottleneck within the conditional behavioral generation process. In our experiments, we demonstrate the effectiveness of our method across various environments in both simulation and the real world.

搜集汇总

背景与挑战

背景概述

CMU Play Fusion是一个专注于机器人技能学习的数据集，利用扩散模型处理非结构化游戏数据的复杂性和多模态性，包含模拟和现实环境数据，用于研究目标导向的技能策略学习。