From backflips to folding laundry: How X Square Robot is building the missing ‘brain’ for embodied AI

A Chinese robotics company has open-sourced three technologies designed to give robots artificial intelligence that works in real-world environments like homes and offices, rather than just performing choreographed tasks. While robot hardware like humanoid bodies and dexterous hands has advanced significantly, the company identifies artificial intelligence as the remaining bottleneck. The three released technologies include a Vision-Language-Action model that converts robot actions into learnable tokens, a World Action Model that teaches robots to understand physical cause and effect rather than just repeat behaviors, and a framework for collecting training data without requiring expensive physical robot demonstrations. Together these tools aim to address core challenges in embodied AI by enabling robots to understand and adapt to the physical world rather than simply execute predetermined movements.