QiXun.AI Spirit v1.5 has soared to the pinnacle of the latest rankings on the embodied intelligence model evaluation platform, RoboChallenge, outperforming Pi 0.5. At present, specific elements of the model, including its weights, have been released as open-source. Subsequent updates are set to be introduced in a phased manner. Spirit v1.5 stands as an end-to-end embodied intelligence foundational model tailored for real-world tasks. It is distinguished by its remarkable capacity for generalization, stability, and accuracy.

This model leverages a unified Vision-Language-Action (VLA) modeling framework. This approach not only minimizes information loss but also bolsters stability. When it comes to training, the model does not hinge on 'pristine' demonstration data. Instead, during the pre-training phase, it adopts open and varied data collection techniques. This enables the model to seamlessly adapt to the intricacies of the real world.

Findings from ablation experiments reveal that models pre-trained on diverse datasets demonstrate superior transfer efficiency and demand fewer computational resources. RoboChallenge, a newly instituted standardized evaluation system in 2025, is dedicated to assessing the cross-platform capabilities of embodied intelligence models. Its evaluation tasks span multiple dimensions, providing a comprehensive assessment framework.