Austin BUDS
收藏资源简介:
BUDS Dataset(Bottom-Up Discovery of sensorimotor Skills)是由德克萨斯大学奥斯汀分校机器人感知与学习小组(RPL)创建的一个用于机器人操作的数据集,旨在通过无分割演示发现可复用的技能,以解决长时序机器人操作任务。该数据集包含来自模拟环境和真实世界任务的多任务演示数据,涵盖工具使用、厨房操作等多种复杂任务。数据集中的演示数据由人类通过遥操作收集,包含多视角图像、机器人本体感知数据和动作指令。数据集的创建过程基于层次聚类算法,通过分析演示数据中的多模态特征,自动将演示序列分割为多个技能片段,并从中提取可复用的技能。这些技能被建模为目标条件的传感器运动策略,能够根据当前状态和目标生成动作指令。数据集的创建无需额外的时间标签,大大减少了人工标注的工作量。该数据集的应用领域主要集中在机器人操作任务,特别是那些需要长时序规划和多阶段操作的复杂场景。BUDS 方法通过发现可复用的技能,显著提高了机器人在多阶段任务中的成功率,平均比现有方法高出 20% 以上。此外,从多任务演示中发现的技能在新任务变体中的平均成功率比单独任务中发现的技能高出 8%。
BUDS Dataset (Bottom-Up Discovery of sensorimotor Skills) was developed by the Robot Perception and Learning (RPL) group at The University of Texas at Austin. It is a robotic manipulation dataset designed to discover reusable skills through unsegmented demonstrations, aiming to solve long-horizon robotic manipulation tasks. This dataset encompasses multi-task demonstration data from both simulated environments and real-world tasks, covering complex scenarios including tool use and kitchen manipulation. The demonstration data is collected by human operators via teleoperation, and includes multi-view images, robot proprioceptive data, and action commands. The dataset construction process leverages hierarchical clustering algorithms, which automatically split demonstration sequences into multiple skill segments and extract reusable skills by analyzing multimodal features within the demonstration data. These skills are modeled as goal-conditioned sensorimotor policies that can generate action commands based on current states and target goals. Notably, the dataset creation eliminates the need for additional time labels, drastically reducing manual annotation workload. The primary application scope of this dataset centers on robotic manipulation tasks, particularly complex scenarios requiring long-horizon planning and multi-stage operations. The BUDS method has significantly improved the success rate of robots in multi-stage tasks, with an average improvement of over 20% compared to existing approaches. Furthermore, skills discovered from multi-task demonstrations achieve an 8% higher average success rate on novel task variants than skills discovered from single-task settings.




