ManiFlow-110k
收藏资源简介:
ManiFlow-110k是一个大规模的3D光学流数据集,包含110000个实例,用于预训练。该数据集通过一个移动对象自动检测流程合成,旨在从人类和机器人操作数据中学习3D流世界模型。该模型预测交互对象在3D空间中的未来运动,指导操作动作规划。数据集内容丰富,包含不同对象、背景、机器人形态和任务,能够有效解决机器人操作技能学习中的挑战,推动该领域的发展。
ManiFlow-110k is a large-scale 3D optical flow dataset containing 110,000 instances for pre-training. This dataset is synthesized via an automated moving object detection pipeline, and is designed to learn 3D flow world models from human and robotic manipulation data. The model predicts the future motion of interacting objects in 3D space to guide manipulation action planning. The dataset covers rich content including diverse objects, backgrounds, robot morphologies and tasks, which can effectively address the challenges in robotic manipulation skill learning and promote the development of this field.

- 13DFlowAction: Learning Cross-Embodiment Manipulation from 3D Flow World Model华南理工大学,腾讯机器人X,香港科技大学,琶洲实验室 · 2025年



