Language-Table 机器人语言标签轨迹数据集

Language-Table 数据集是 Google 于 2022 年提出的，用于构建能够在现实世界中通过自然语言指令进行实时互动的机器人。相关论文成果为「Interactive Language: Talking to Robots in Real Time」。该数据集包含了近 600,000 条带有语言标签的轨迹，用于推动更先进、更有能力、可自然语言交互的机器人发展。通过在包含数十万条带有语言注释的轨迹数据集上进行训练，研究人员发现所得的策略能够执行比以往多 10 倍的指令，这些指令描述了现实世界中的端到端视听觉-运动技能。该数据集包含了近 600,000 条带有语言标签的轨迹, 用于推动更先进、更有能力、可自然语言交互的机器人发展。

The Language-Table dataset was proposed by Google in 2022 for developing robots that can interact in real time via natural language instructions in the real world. Its associated research paper is titled "Interactive Language: Talking to Robots in Real Time". This dataset contains nearly 600,000 language-labeled trajectories, aiming to advance the development of more advanced, capable robots that support natural language interaction. After training on this large-scale trajectory dataset with hundreds of thousands of language annotations, researchers found that the resultant policies can execute 10 times more instructions than previous works, where these instructions describe end-to-end audiovisual-motor skills in the real world.

搜集汇总

背景与挑战

背景概述

Language-Table是Google在2022年提出的机器人数据集，包含约60万条带语言标签的轨迹，旨在推动机器人通过自然语言指令进行实时交互。该数据集用于训练端到端视听觉-运动技能，使机器人能执行更广泛的现实世界指令。