XRZero-G0: A framework for high-quality robot-free data collection and embodied AI training

SHENZHEN, China, June 10, 2026 /PRNewswire/ -- Scaling embodied AI has long been bottlenecked by data. Teleoperating real robots is expensive and slow, yielding only a limited number of demonstrations per day. While robot-free data collection offers a promising alternative, the lack of systematic quality control and training integration has limited its effectiveness for policy learning.

X Square Robot announces the open-source release of XRZero-G0, a hardware-software co-designed framework for robot-free data collection, trainable policy generation, and real-robot evaluation. Alongside it, the team releases G0-Dataset, a large-scale validated multimodal dataset produced by XRZero-G0, providing reproducible high-quality robot-free data for the global robotics community.

Bridging robot-free and real-world perception

Physical robots perceive the world through multiple viewpoints, typically a head-mounted camera for global context and wrist-mounted cameras for fine-grained manipulation. In contrast, most robot-free systems rely only on wrist-view observations from human demonstrators, creating a gap between training and deployment.

XRZero-G0 addresses this gap with a multi-view aligned sensing system that aligns human demonstration with robot observation spaces.

The system combines a head-mounted camera and dual wrist cameras to capture both global context and detailed hand-object interactions. These synchronized observations are mapped into a shared representation compatible with robot perception.

A wearable VR interface and interchangeable grippers allow human operators to generate demonstrations that are directly transferable to different robot embodiments, enabling high-throughput robot-free data collection across diverse environments.

Making robot-free demonstrations truly trainable

Data quality has been a critical barrier in robot-free learning. XRZero-G0 formalizes trainability governance via a closed-loop Collection–Inspection–Training–Evaluation pipeline:

• Observation level: multi-view geometric consistency suppresses visual-kinematic misalignment.
• Kinematic level: full-body inverse kinematics with collision and joint-limit constraints filters invalid trajectories.
• Policy level: real-robot playback serves as the final validation criterion.

This pipeline improves the usability of robot-free demonstrations, with experiments showing an effective data yield of around 85% under controlled experimental settings, significantly increasing the proportion of trainable samples.

A 10:1 mixing law reduces real-robot data requirements

A key finding of the XRZero-G0 study is that robot-free data and real-robot data can complement each other effectively.

Controlled experiments show that combining approximately 10 robot-free episodes with 1 real-robot episode achieves performance comparable to purely real-robot datasets in evaluated tasks.

Robot-free data provides broad behavioral coverage and task understanding, while a small amount of real-robot data anchors embodiment-specific factors such as motor latency and friction. This strategy reduces the need for real-robot data by up to 20× under experimental conditions.

G0-Dataset scales XRZero-G0 into a 2,000-hour dataset

Built on XRZero-G0, G0-Dataset provides over 2,000 hours of validated multimodal demonstrations spanning vision, tactile, and audio modalities.

The dataset integrates robot-free collection, automated quality inspection, mixed-data training, and real-robot evaluation for research purposes. G0-Dataset supports large-scale pretraining and cross-embodiment transfer experiments, providing a reproducible open resource for robotics research.

Zero-shot transfer across robot embodiments

Experiments indicate that policies trained with XRZero-G0 exhibit improved generalization across collection environments, including varying robot poses, table heights, and viewpoints.

They also demonstrate zero-shot cross-embodiment transfer ability in evaluated settings, where policies trained with mixed data can be transferred to unseen robot platforms without task-specific fine-tuning.

Building an open ecosystem

By open-sourcing XRZero-G0 and releasing G0-Dataset, X Square Robot provides hardware designs, automated inspection pipelines, training methodologies, and high-quality datasets to the research community.

These resources aim to accelerate the development of general-purpose robots and scalable embodied AI, supporting a transition toward more systematic and large-scale data generation approaches.

XRZero-G0 and G0-Dataset are now publicly available for researchers and developers worldwide.

Project Homepage: https://x2robot.com/x2go
Paper: https://arxiv.org/abs/2604.13001
Code: https://github.com/X-Square-Robot/XRZero-G0
Open Dataset: https://huggingface.co/datasets/x-square-robot/XRZero-G0-3K 

Media Inquiries: contact@x2robot.com