MoCapAct

MoCapAct是一个用于模拟类人控制的多任务数据集，由乔治亚理工学院机器人与智能机器研究所和微软研究院合作创建。该数据集包含了专家代理及其轨迹，这些代理能够跟踪超过三小时的模拟类人在dm_control物理环境中的运动捕捉数据。数据集中的专家代理经过训练，可以跟踪来自CMU运动捕捉数据集的记录运动，包括站立、行走、跑步等基本动作。MoCapAct数据集的创建旨在降低基于学习的类人控制研究的门槛，通过提供高质量的代理和轨迹数据，支持训练能够执行多种动作的单一层次策略，并展示这些策略在学习和执行高级任务中的效率。此外，数据集还被用于训练自回归GPT模型，以实现在给定运动提示下的自然运动完成。

MoCapAct is a multi-task dataset for humanoid control simulation, co-created by the Institute for Robotics and Intelligent Machines at Georgia Institute of Technology and Microsoft Research. The dataset comprises expert agents and their corresponding trajectories, which are capable of tracking over three hours of motion capture data for simulated humanoids within the dm_control physics simulation environment. The expert agents in the dataset are trained to track recorded motions sourced from the CMU Motion Capture Dataset, including basic movements such as standing, walking, running, and more. The MoCapAct dataset was developed to lower the barrier to entry for learning-based humanoid control research. By providing high-quality agent and trajectory data, it supports the training of a single hierarchical policy that can execute diverse actions, and demonstrates the efficiency of such policies in learning and executing advanced tasks. Furthermore, the dataset has been utilized to train autoregressive GPT models to enable natural motion completion given motion prompts.

搜集汇总

背景与挑战

背景概述

MoCapAct是一个由乔治亚理工学院和微软研究院合作创建的多任务数据集，专注于模拟类人控制。它包含基于CMU运动捕捉数据的专家代理和超过三小时的模拟轨迹，覆盖站立、行走、跑步等基本动作，旨在通过提供高质量数据降低类人控制研究门槛，支持训练单一层次策略和自回归GPT模型，以高效执行高级任务和自然运动完成。