X-Square-Robot-Open-Sources-Autonomous-Data-Collection-Framework-for-Free-Use,

Post Views: 8

Companies developing physical robots face significant challenges in gathering sufficient training data through traditional methods. The process of manually operating robots to collect demonstrations is time-consuming and costly, producing limited datasets that hinder the advancement of embodied AI systems.

XRZero-G0: Hardware and Software Framework

X Square Robot has introduced XRZero-G0, a hardware and software framework designed to collect training data using human operators rather than physical robots. The system, released under an MIT license, includes G0-Dataset, a multimodal dataset created through this approach.

Multi-Camera Setup

The system employs a multi-camera setup to align human demonstrations with robot perception. A head-mounted camera captures environmental context, while wrist-mounted cameras track detailed hand and object interactions. This dual-camera configuration ensures comprehensive data collection that mirrors how robots interpret tasks during deployment.

Unified Representation and VR Interface

The framework integrates synchronized visual inputs into a unified representation compatible with robotic sensing. A VR interface and modular grippers enable operators to generate demonstrations applicable across various robot configurations.

Data Validation Process

Data quality is prioritized through a closed-loop validation process. At the observation level, multi-view geometric consistency minimizes image and motion misalignment. Kinematic checks using whole-body inverse kinematics with collision and joint-limit constraints eliminate invalid movement trajectories. Final validation occurs through physical robot playback, ensuring reliability.

Testing and Efficiency

Testing under controlled conditions achieved an 85% effective data yield. The framework demonstrates a 10-to-1 data efficiency ratio, where 10 human-collected episodes combined with one real-robot episode achieve performance comparable to datasets composed entirely of robot-generated data.

G0-Dataset: Multimodal Training Data

G0-Dataset contains over 2,000 hours of validated demonstrations spanning vision, tactile, and audio modalities. It encompasses 3,000 distinct manipulation tasks, ranging from basic operations to complex semantic actions, following a long-tail distribution. Operators achieved a peak collection rate of 93.2 episodes per hour.

Generalization and Cross-Embodiment Research

The dataset supports large-scale pretraining and cross-embodiment research. Policies trained using the framework exhibit generalization across varying robot postures, table heights, and viewpoints, as well as zero-shot transfer to untrained robot platforms without platform-specific adjustments.

Significance of the Release

The release of XRZero-G0 and G0-Dataset represents a significant advancement in robot training data collection, offering a scalable solution that balances cost efficiency with technical precision. By leveraging human operators while maintaining rigorous quality controls, the framework addresses critical bottlenecks in embodied AI development.

Human-collected data provides broad behavioral diversity, while real-robot data incorporates physical constraints like motor latency and friction. This approach reduces real-robot data requirements by up to 20 times in tested scenarios.