OXE-AugE: A Large-Scale Robot Augmentation of OXE for Scaling Cross-Embodiment Policy Learning

📝 Original Info

• Title: OXE-AugE: A Large-Scale Robot Augmentation of OXE for Scaling Cross-Embodiment Policy Learning
• ArXiv ID: 2512.13100
• Date: 2025-12-15
• Authors: Guanhua Ji, Harsha Polavaram, Lawrence Yunliang Chen, Sandeep Bajamahal, Zehan Ma, Simeon Adebola, Chenfeng Xu, Ken Goldberg

📝 Abstract

Figure 1: We present OXE-AugE, a large-scale open-source dataset that augments the Open-X Embodiment (OXE) dataset [1] with 9 different robot embodiments across 16 datasets, covering 60% of the widely-used Octo pretraining mixture [2]. In total, OXE-AugE provides over 4 million trajectories, more than triple those in the original OXE. Robots in OXE-AugE include Panda, UR5e, Xarm7, Google robot, widowX, Sawyer, Kinova3, IIWA, and Jaco. We find that training on OXE-AugE improves OpenVLA [3] and 𝜋 0 [4] policy performance by up to 24-45% on previously unseen robot-gripper combinations across four real-world manipulation tasks. Large and diverse datasets are needed for training generalist robot policies that have potential to control a variety of robot embodiments-robot arm and gripper combinations-across diverse tasks and environments. As re-collecting demonstrations and retraining for each new hardware platform are prohibitively costly, we show that existing robot data can be augmented for transfer and generalization. The Open X-Embodiment (OXE) dataset, which aggregates demonstrations from over 60 robot datasets, has been widely used as...

📄 Full Content

...(본문 내용이 길어 생략되었습니다. 사이트에서 전문을 확인해 주세요.)