Zi变量 Unveils WALL-B Embodied AI Model; Robots to Enter Real Homes in 35 Days

Gate News message, April 21 — Zibianliang (自变量), a Chinese robotics company, held a press conference on April 21 to unveil its next-generation embodied AI foundation model, WALL-B. The company announced that robots powered by WALL-B will enter real households in 35 days.

According to Zibianliang co-founder and CTO Wang Hao, WALL-B is built on a World Unified Model (WUM) architecture, designed to eliminate data loss between separate modules. Unlike traditional vision-language-action (VLA) models where visual, language, and motion modules operate independently—causing information loss with each data transfer—WALL-B integrates vision, language, action, and physical prediction capabilities into a single unified network trained jointly from scratch. Wang emphasized that world models are not separate plug-in modules, but rather predictive capabilities for the physical world's future states.

The company's core insight centers on data quality: Wang Hao distinguished between "sugar water data" (clean, stable, predictable lab data) and "milk data" (messy, uncontrollable, real-world household data). While training on lab data produces models lacking zero-shot generalization, real household data—though costly and time-consuming to collect—enables true generalization. To this end, Zibianliang has entered over 100 volunteer homes to train WALL-B.

CEO Wang Qian stated that robots can perform any physically feasible task once deployed in homes, requiring no advance consideration of limitations. He highlighted that competitive advantage stems not from algorithms or hardware, but from the complete engineering ecosystem—data definition, collection, processing, and training evaluation. In the robotics field, such technological leadership windows could extend three years or longer. Notably, Zibianliang recently completed its Series B funding round led by Xiaomi's venture arm, bringing the company's disclosed backers to four major Chinese internet firms (ByteDance, Meituan, Alibaba, and Xiaomi).