At WAIC 2026, MagicLab's Magic-VLA K02 model demonstrated three high-difficulty, long-horizon household tasks: box folding and sealing, flexible garment organizing, and suitcase packing — a deliberately chosen set that stresses four distinct capabilities at once: rigid-object manipulation, deformable-material handling, multi-object spatial reasoning, and task-interruption recovery.
The standout result: K02's integrated box-folding-and-sealing long-horizon task — an industry first for a general-purpose model attempting this specific task chain — achieved a success rate above 90%, a benchmark few general-purpose (rather than single-task-trained) models have publicly reported for a task this long and multi-step.
Why it matters for smart-home and robotics-as-a-service buyers: Most household robot demos show a single short action — picking up one object, folding one item. A long-horizon task chain (fold, then seal, in sequence, on varied box sizes) with a 90%+ success rate is a much stronger indicator of real household usefulness than a series of isolated single-action demos strung together.
English 
