Open X-Embodiment: Unified Robotic Policy Learning

• OXE is a unified initiative that consolidates heterogeneous robotic demonstrations across multiple tasks and embodiments for training flexible, generalist policies.
• It aggregates over 1.4M trajectories from 22 robot types and 527 skill categories using standardized formats and transformer-based methods to achieve cross-embodiment generalization.
• OXE employs augmentation techniques such as synthetic data generation and mask fusion to enhance data balance, reduce engineering overhead, and improve sample efficiency.

Open X-Embodiment (OXE) is a coordinated initiative, dataset, and modeling paradigm designed to enable generalist robot policies through standardized, large-scale pooling of demonstration data across heterogeneous robot platforms, tasks, and environments. By aggregating manipulation datasets from over 60 sources, encompassing a wide variety of robot embodiments and tasks, OXE provides the infrastructure and reference methods to train transformer-based policies that generalize across hardware. The project’s impact spans manipulation, locomotion, and navigation, with OXE-derived methodologies and augmentations now widely adopted in cross-embodiment vision-language-action learning.

1. Motivation and Conceptual Foundations

OXE arises from the fragmentation in traditional robotic learning, where models are typically trained per-robot, per-task, and per-environment, leading to duplicated effort and narrow generalization. The motivation is to provide a consolidated, standardized, and open dataset that enables large-capacity models to exploit shared structure across diverse robots similarly to the role played by foundation models in vision and NLP. The intended benefits of such "X-robot" policies are: (a) positive transfer, where models improve by leveraging multiple robots’ experiences for similar skills; (b) improved sample efficiency, especially for hardware with limited data; (c) reduced engineering overhead, enabling fine-tuning or direct deployment of a single policy to new embodiments and tasks (Collaboration et al., 2023).

2. OXE Dataset: Structure, Composition, and Augmentation

The original OXE dataset aggregates over 1.4 million successful real-robot trajectories across 22 robot embodiments (single-arm, bimanual, quadruped manipulators) and over 527 skill categories in 160,266 unique task instances, standardized via RLDS and TFRecord formats. Each episode records:

• Observations: Fixed camera RGB (or RGB-D/point cloud), task metadata (language instruction, scene ID).
• Actions: 7-DoF end-effector command (position, orientation, gripper), normalized per-dataset and discretized to 256 bins/dimension.
• Metadata: Robot type, coordinate frame, skill tags, institution (Collaboration et al., 2023).

Skill tagging is performed using an LLM (PaLM), enabling a systematic taxonomy for tasks (pick–place, assembly, wipe, insert, etc.) and object types. Heterogeneity is addressed by canonicalizing action and observation spaces and allowing coordinate frame differences to be implicitly inferred.

OXE-AugE addresses severe dataset imbalance (top 4 robots account for >85% of real data) by synthetically augmenting source trajectories into multiple target robot embodiments using AugE-Toolkit. This includes mask fusion (geometric and learned segmentations), automatic base-pose tuning, and retargeted simulation replay to generate additional trajectories in new embodiments, tripling the scale (4.44M trajectories, 9 robot–gripper types) and achieving near-uniform embodiment entropy $H \approx 0.97$ (Ji et al., 15 Dec 2025). This augmentation systematically improves generalization to both seen and unseen robots.

3. Architectures for Cross-Embodiment Policy Learning

**RT-X Models (RT-1