libero_trajectory_augmented
收藏资源简介:
Libero Trajectory Augmented Dataset是一个用于机器人学习的增强轨迹数据集,专门为LIBERO基准测试设计。该数据集旨在解决离线强化学习和视觉-语言-动作模型训练中专家演示分布狭窄的问题,通过提供分布外恢复轨迹来人工增强离线数据集的鲁棒性,无需人工遥操作。数据生成采用逆向动作范式:在原始专家演示的关键状态(如抓取或交互前)注入有界6D坐标噪声到机器人末端执行器状态,迫使专家策略从噪声干扰的分布外状态恢复并完成任务。每个原始专家演示生成2个增强轨迹,使数据集规模增至原始的三倍。这种噪声注入创造了轨迹长度的变异性,迫使模型学习恢复动作,提高了分布外鲁棒性。数据集采用原生Hugging Face LeRobot v3.0格式,包含高度压缩的Parquet文件用于表格化状态/动作数据,以及高效SVT-AV1分块MP4视频用于视觉观察。方向为完美原生OpenCV格式(正立、未镜像),已动态校正底层Robosuite OpenGL渲染伪影以匹配8D NVIDIA模式。数据集包含三个LIBERO套件的完整处理子目录:libero_goal、libero_spatial和libero_object。适用于训练需要从分布外状态恢复的鲁棒机器人策略,特别是需要处理轨迹长度变异性和空间配置变化的场景。
Libero Trajectory Augmented Dataset is a trajectory-augmented dataset for robotic learning, specifically designed for the LIBERO benchmark. This dataset aims to address the issue of narrow expert demonstration distributions in offline reinforcement learning and vision-language-action model training, artificially enhancing the robustness of offline datasets by providing out-of-distribution recovery trajectories without requiring manual teleoperation. The data generation adopts the reverse action paradigm: bounded 6D coordinate noise is injected into the robot end-effector state at key states of the original expert demonstrations (e.g., before grasping or interaction), forcing the expert policy to recover from the noise-disturbed out-of-distribution states and complete the task. Each original expert demonstration generates 2 augmented trajectories, increasing the dataset size to three times the original scale. This noise injection creates variability in trajectory length, forcing models to learn recovery actions and improving out-of-distribution robustness. The dataset uses the native Hugging Face LeRobot v3.0 format, including highly compressed Parquet files for tabulated state/action data, and efficient SVT-AV1 chunked MP4 videos for visual observations. The orientation follows the perfect native OpenCV format (upright, not mirrored), and underlying Robosuite OpenGL rendering artifacts have been dynamically corrected to match the 8D NVIDIA mode. The dataset includes fully processed subdirectories for the three LIBERO suites: libero_goal, libero_spatial, and libero_object. It is suitable for training robust robotic policies that require recovery from out-of-distribution states, especially scenarios that need to handle trajectory length variability and spatial configuration changes.
数据集名称
Libero Trajectory Augmented Dataset
许可证
Apache-2.0
数据集概述
该数据集为 LIBERO 基准测试的官方格式 LeRobot (v3.0+) 数据集,通过物理状态扰动(Physical State Perturbation)进行了增强。旨在通过提供分布外(OOD)恢复轨迹来增强离线数据集的鲁棒性,无需人工遥操作。通过学习这些增强轨迹,策略能够接触到更高的轨迹长度变异性、改变的空间配置,并学习明确的恢复行为。
增强方法
- 生成引擎:使用 Robosuite MuJoCo 引擎动态生成轨迹。
- 方法:采用逆动作范式,具体步骤包括:
- 加载原始专家演示。
- 将环境推演到关键状态(例如,抓取或交互前的时刻)。
- 向机器人末端执行器状态注入有界的 6D 坐标噪声。
- 强制专家策略从噪声、分布外的状态中恢复并成功完成任务。
数据集规格
- 增强规模:每个原始专家演示生成
2条增强轨迹,有效将原始数据集大小扩大 三倍。 - 变异性:噪声注入带来轨迹长度变异性,迫使模型学习恢复动作,提升分布外鲁棒性。
- 方向:完美原生 OpenCV(正立、无镜像)。生成过程中动态校正了底层 Robosuite OpenGL 渲染伪影(X/Y 轴反转),以匹配 8D NVIDIA 原生模式。
- 格式:原生 Hugging Face LeRobot
v3.0数据集,包含:- 高度压缩的 Parquet 文件存储表格状态/动作数据。
- 高效
SVT-AV1分块 MP4 视频存储视觉观察。
包含的数据子集(Suites)
本仓库包含以下 LIBERO 子目录的完整处理数据:
libero_goallibero_spatiallibero_object
使用示例
使用 Hugging Face LeRobotDataset API 加载数据:
python
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
加载增强后的 libero_goal 数据集
dataset = LeRobotDataset("JDhruvR/libero_trajectory_augmented/libero_goal")





