Mitsubishi Electric's MEL-Craft wins first Humanoid IKEA Assembly Challenge

The challenge itself was new: IROS 2026 in Pittsburgh hosted the first Humanoid IKEA Assembly Challenge, a contest built around a common humanoid robot platform supplied to every team. That shared-hardware setup makes the comparison a test of software and control rather than of mechanical design, which is why the result is framed as a measure of physical AI capability.

The task was deliberately narrow but fiddly. Teams had to get a humanoid to recognize parts, grasp them, and fit four legs onto the tabletop of IKEA's UTTER children's table, continuing without human help. Scoring rewarded progress through the assembly steps and speed, while human assistance counted against a team.

MEL-Craft's approach paired recognition modules with a Vision-Language-Action model — an AI model that turns visual and language input into robot motion — together with lower-body and upper-body control. Mitsubishi Electric frames the know-how from recognizing and gripping parts through to fitting them as relevant to automating assembly work in manufacturing, and says it will keep pursuing physical AI research using what the team learned.