Open X-Embodiment Alternative: Targeted Commercial Data for Production Robotics

Open X-Embodiment is the single most important dataset for pretraining generalist robot policies, and its cross-embodiment diversity enables the kind of broad manipulation priors that single-robot datasets cannot provide. However, the transition from generalist pretraining to production deployment reveals OXE's structural limitations: mixed licensing creates legal risk, inconsistent quality introduces training noise, and the data inevitably mismatches your specific robot platform and deployment environment. Claru addresses each of these gaps. We provide commercially licensed, quality-controlled demonstrations collected on your exact robot, in your actual facility, with standardized action spaces and full multi-modal sensor coverage. Teams that pretrain on OXE and fine-tune on Claru data consistently achieve higher success rates than either source alone, because the combination captures OXE's breadth for general manipulation reasoning while grounding policy behavior in the real-world specifics of the deployment domain. We deliver in RLDS format for seamless integration with OXE-based training pipelines, and our ongoing collection service means your fine-tuning dataset grows alongside your deployment requirements rather than remaining static.

Open X-Embodiment (OXE) is a collaborative dataset and model initiative led by Google DeepMind in partnership with 21 academic institutions worldwide. Published in 2023 and presented at ICRA 2024, it represents the most ambitious effort to date to unify robot learning data across diverse embodiments. The dataset aggregates over 1 million robot demonstrations from 22 distinct robot platforms -- including the Google Robot, WidowX Bridge, Franka, Kuka, xArm, Tiago, and others -- spanning over 527 skills across 160,000+ task configurations.

The core motivation behind OXE was to enable generalist robot policies that transfer across embodiments. The project demonstrated this with RT-1-X and RT-2-X, showing that policies trained on the combined OXE dataset outperformed those trained on any single constituent dataset on most platforms. Each contributing dataset was converted to the RLDS (Reinforcement Learning Datasets) format hosted on TensorFlow Datasets, establishing RLDS as a de facto standard for robot learning data interchange.

OXE's constituent datasets vary enormously in character. Bridge Data V2 contributes ~60,000 demonstrations of kitchen manipulation with a WidowX arm. RT-1 contributes ~130,000 demonstrations from Google's mobile manipulators. Fractal and Kuka contribute industrial arm data. Some datasets include language annotations; others do not. Some record at 5 Hz; others at 30 Hz. Action spaces range from 2-DoF end-effector deltas to 7-DoF joint velocities. This heterogeneity is both OXE's strength (diversity) and its challenge (inconsistency).

The dataset is released under mixed licensing that varies per constituent dataset. Some subsets like Bridge V2 use permissive licenses, while others carry non-commercial restrictions. OXE has become the standard pretraining corpus for generalist robot policies including Octo, OpenVLA, and CrossFormer, and its RLDS format has been widely adopted by the robot learning community.

OXE's most pressing production challenge is licensing. Because the dataset aggregates contributions from 21 institutions, each subset carries its own license. Several subsets restrict commercial use, and the licensing status of some contributions is ambiguous. Production teams must audit every subset they train on to confirm commercial rights -- a legal burden that grows with each new constituent added to the collection. There is no single 'OXE commercial license' that covers the full dataset.

Quality consistency is a systemic issue. OXE's constituent datasets were collected by different labs, with different teleoperation systems, different data quality standards, and different annotation practices. Some datasets contain carefully curated demonstrations with language annotations; others include noisy trajectories or failed attempts with no language labels. The RT-1 subset is known for high quality, while some smaller contributions have minimal curation. Training on the full mix without filtering often introduces noise that degrades performance on specific deployment domains.

Action space heterogeneity creates significant engineering challenges. OXE datasets use different action representations -- some record end-effector position deltas, others record joint velocities, and some use absolute poses. Action dimensions vary from 2-DoF (planar pushing) to 7-DoF (full arm) to 8-DoF (arm + gripper). Normalizing these into a common space for cross-embodiment training is a non-trivial preprocessing step that every user must implement, and the normalization choices themselves affect policy performance.

Temporal resolution varies from 3 Hz to 30 Hz across subsets, meaning that a 'step' in one dataset represents a fundamentally different time quantum than in another. This temporal inconsistency complicates training of action-chunking architectures that assume a fixed control frequency.

Finally, OXE's environments are overwhelmingly research labs. The visual backgrounds, lighting conditions, and scene compositions reflect academic settings, not the production environments (factories, warehouses, commercial kitchens, retail floors) where robots actually deploy. Policies pretrained on OXE still require substantial environment-specific data to generalize to real deployment conditions.

OXE is the clear choice for pretraining generalist robot policies. Its cross-embodiment diversity teaches policies to extract embodiment-agnostic representations -- spatial reasoning, object affordances, and manipulation primitives that transfer across robot morphologies. This is precisely why Octo, OpenVLA, and similar foundation models use OXE as their pretraining corpus. If you are building or fine-tuning a generalist policy, starting from an OXE-pretrained checkpoint gives you a significant head start over training from scratch.

OXE is also valuable for cross-embodiment research. If your research question involves transfer learning across robot platforms, few-shot adaptation to new embodiments, or the relationship between pretraining diversity and downstream performance, OXE provides the multi-embodiment data you need for rigorous experiments.

Switch to commercial data when you need production reliability on a specific deployment. OXE's broad diversity is a liability when your goal is high success rates on a narrow task set. A policy trained on 1M demonstrations from 22 robots does not necessarily outperform one fine-tuned on 10K high-quality demonstrations from your specific robot doing your specific tasks. Claru provides the targeted, high-quality, domain-specific data that transforms a generalist pretrained model into a production-ready policy.

The predominant industry pattern is OXE for pretraining, then commercial fine-tuning data for deployment. This combination captures OXE's breadth for general capability while using Claru's depth for deployment-specific performance.

Claru provides the fine-tuning data layer that makes OXE-pretrained policies production-ready. Where OXE gives you breadth across 22 embodiments, Claru gives you depth on your single deployment target. Our demonstrations are collected by trained teleoperators on your physical robot, in your facility, manipulating your actual objects -- eliminating the domain gap between research-lab data and production conditions.

Unlike OXE's heterogeneous quality, every Claru demonstration passes a multi-stage quality pipeline. Automated checks flag kinematic anomalies, excessive force events, and out-of-workspace excursions. Human reviewers verify task completion and validate language annotations against the demonstrated behavior. The result is a uniformly high-quality dataset where every trajectory represents a successful, well-executed demonstration of the target task.

Claru data comes with a single, clear commercial license that covers all delivered demonstrations. There is no need to audit subsets or navigate mixed licensing terms. You receive full IP rights to the data, suitable for training models deployed in commercial products.

We deliver in RLDS format (matching OXE's standard) as well as HDF5, zarr, and LeRobot. The action space, observation space, and control frequency are standardized to your robot's specifications from the start -- no normalization pipeline required. Claru data can be mixed directly with OXE subsets for co-training or used independently for fine-tuning on a pretrained checkpoint.