OpenDriveLab/EgoHumanoid
收藏资源简介:
EgoHumanoid样本数据集(G1)是一个与论文《EgoHumanoid: Unlocking In-the-Wild Loco-Manipulation with Robot-Free Egocentric Demonstration》同时发布的样本数据集,用于快速实验和微调。它包含两种数据源:机器人遥操作演示(使用Unitree G1机器人)和机器人无关的以自我为中心的人类演示(使用PICO VR和ZED设备)。数据集是一个样本子集,其中机器人子集包含50个演示片段,人类子集包含1个演示片段。该数据集旨在用于测试EgoHumanoid协同训练流程(结合机器人和人类数据)、在小型G1任务上微调预训练的VLA策略,以及验证数据加载器和LeRobot兼容的存储格式。数据集结构遵循LeRobot v2约定,便于兼容LeRobot加载器。
The EgoHumanoid Dataset (G1) is a sample dataset released concurrently with the paper *EgoHumanoid: Unlocking In-the-Wild Loco-Manipulation with Robot-Free Egocentric Demonstration*, intended for rapid experimentation and fine-tuning. It comprises two types of data sources: robot teleoperation demonstrations (using the Unitree G1 robot) and robot-free egocentric human demonstrations (using PICO VR and ZED devices). This dataset is a sample subset, where the robot subset contains 50 demonstration segments and the human subset contains 1 demonstration segment. This dataset is designed for testing the EgoHumanoid collaborative training pipeline (combining robot and human data), fine-tuning pre-trained VLA policies on small-scale G1 tasks, and verifying data loaders and LeRobot-compatible storage formats. The dataset structure follows the LeRobot v2 conventions to ensure compatibility with LeRobot loaders.




