OXE-AugE

OXE-AugE is an open-access synthetically augmented robotics dataset developed by UC Berkeley's AUTOLab and collaborators to scale cross-embodiment generalist robot policy learning. Created using the simulation-assisted 2D cross-painting pipeline AugE-Toolkit, the collection systematically expands 16 core real-world manipulation datasets from the Open X-Embodiment collection into over 4.4 million trajectories spanning up to 9 distinct robot arm and gripper combinations. Source demonstrations primarily featuring platforms such as Franka Panda, Universal Robots UR5e, xArm, and Google Robot are visually projected onto Sawyer, Kinova Gen3, and KUKA iiwa embodiments to remediate pervasive data imbalances. The dataset targets robotic foundation model researchers training vision-language-action architectures like OpenVLA and pi0 on diverse quasi-static manipulation and tabletop pick-and-place tasks without requiring costly physical data recollection.

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

Overview

OXE-AugE is an open-source, large-scale robot embodiment augmentation dataset developed by researchers at UC Berkeley's AUTOLab in affiliation with the Berkeley AI Research (BAIR) Lab. Built on top of the Open X-Embodiment (OXE) corpus, OXE-AugE addresses hardware imbalance in robotics datasets, where over 85% of real demonstrations originate from just four robot platforms.

Methodology & AugE-Toolkit

OXE-AugE scales cross-embodiment generalization using the AugE-Toolkit, an automated 2D image-space cross-painting pipeline:

• Source Selection: Curates 16 high-utility datasets from the OXE collection (covering roughly 60% of the Octo pretraining mix, including Bridge and RT-1 datasets, totaling approximately 550,000 source trajectories).
• Embodiment Augmentation: Demonstrations collected on original arms (such as Franka Panda, UR5, xArm, Google Robot, WidowX, and Jaco) are augmented across up to 9 different platforms, adding Sawyer, Kinova Gen3, and KUKA iiwa setups.
• Simulation Replay: Uses simulation replays with camera orientation matching to superimpose alternative robot geometries and grippers (e.g., Robotiq 2F-85, Rethink Gripper) directly into existing real-world video frames.
• Dataset Scale: Multiplies source demonstrations into 4.44 million trajectories, more than triple the original OXE manipulation size.

Use Cases

• Pretraining Generalist Robot Policies: Used to pretrain and fine-tune Vision-Language-Action (VLA) foundation models, such as OpenVLA and $\pi_0$.
• Cross-Embodiment Transfer: Significantly enhances zero-shot and few-shot deployment success on novel or previously unseen robot-gripper configurations.

Frequently asked questions

What is OXE-AugE?

OXE-AugE is an open-source augmented robot dataset created by UC Berkeley's AUTOLab that expands 16 core Open X-Embodiment datasets across 9 robot embodiments, scaling the collection to over 4.4 million trajectories.

Which robot embodiments are included in OXE-AugE?

OXE-AugE features 9 robot platforms: Franka Emika Panda, UR5e, xArm7, WidowX, Google Robot, Kinova Jaco, Sawyer, Kinova Gen3, and KUKA iiwa.

How is the dataset generated?

OXE-AugE is produced using AugE-Toolkit, which matches camera parameters in simulation and cross-paints different robot arms and grippers into original demonstration video frames in 2D image space.

What is the primary license for OXE-AugE datasets?

The datasets are distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license, while the accompanying code is released under open-source Apache-2.0 / MIT licenses.

Where can the dataset and code be accessed?

The dataset shards are hosted on Hugging Face (oxe-auge organization), and the tooling code is hosted on GitHub under GuanhuaJi/oxe-aug.