Magic Atom's Magic-VLA K02 Achieves Over 90% Success Rate in Box Stacking and Sealing Long-Horizon Task at WAIC

At the 2026 World Artificial Intelligence Conference (WAIC), held July 17-20 in Shanghai, Magic Atom demonstrated its general-purpose embodied large model Magic-VLA K02 successfully completing long-horizon box stacking and sealing tasks, drawing significant attention from attendees. The company stated this marks the first time a general-purpose embodied model has fully realized the combined box-stacking-and-sealing workflow. Robots equipped with Magic-VLA K02 autonomously handled task understanding, step planning, object recognition, and continuous action execution, while also adjusting strategies in real-time when environmental conditions changed or workflows were interrupted.

Magic-VLA K02 employs a two-tier dual-system architecture. The high-level system handles global decision-making and task planning, decomposing goals into executable atomic tasks and generating key result images as visual target references. The low-level system, composed of a VLM backbone network, dynamic expert modules, and action expert modules, translates the planned atomic tasks into continuous robotic actions.

During on-site testing, Magic-VLA K02 achieved 92% overall accuracy on complex long-horizon tasks, with task interruption rates reduced to under 5%. Trajectory deviation correction response speed improved by 70%, and required robot demonstration training data decreased by 60%. These results demonstrate significant improvements in long-horizon strategic control, skill combination generalization, cross-robot adaptation, and deployment stability.

Beyond box stacking and sealing, the model also performed flexible clothing folding and luggage packing tasks. Clothing folding targets highly deformable object manipulation—garments lack stable geometric forms, yet the model dynamically selected grasp points and adjusted action parameters based on real-time visual feedback. Luggage packing tested the model's ability to handle constraint relationships between multiple objects and limited space.

Box stacking and sealing are fundamental operations in logistics, warehousing, industrial production, and e-commerce retail, and are widely recognized as challenging tasks in embodied AI. They involve fine-grained rigid object manipulation and dynamic deformable object control respectively, with automation levels directly impacting packaging efficiency, operational standardization, and labor costs.

For cross-robot adaptation, Magic-VLA K02 introduces a metadata description system that unifies representation of different robot body types, control modes, and action spaces. This enables the same model framework to support robotic arms, mobile manipulators, and single- or dual-arm configurations. Within tested devices, cross-platform adaptation reached 100% success, with compatible device categories increasing by over 200%.

The WAIC demonstration validates Magic-VLA K02's capabilities in long-horizon planning, deformable object manipulation, dynamic disturbance recovery, and multi-object spatial reasoning, further establishing its value as Magic Atom's core embodied intelligence platform. The company's full product line has been listed on the Wanji Yizu 2.0 rental platform, accelerating deployment into real-world workplaces.

As more embodied AI breakthroughs emerge at WAIC 2026, expectations are rising for general-purpose robots entering industrial production and daily service scenarios. Magic-VLA K02's proven long-horizon task capabilities offer a reference path for multi-form robots transitioning toward scalable deployment.