en.Wedoany.com Reported - On April 21, X Square Robot held a launch event in Beijing, officially unveiling its new self-developed embodied intelligence foundation model, WALL-B. It was announced that the new generation of robots equipped with this model would enter real households in the first batch 35 days later. This is the world's first embodied intelligence foundation model based on a unified world model architecture, marking a significant leap from the VLA architecture to a native multimodal fusion architecture. Independent Variable also announced the recent completion of its Series B funding round, led by Xiaomi Strategic Investment and co-led by Sequoia Capital China, with a total funding scale of nearly 2 billion yuan. Following this round, Independent Variable has become the only embodied intelligence enterprise in China to have received investments from all four major internet giants: ByteDance, Meituan, Alibaba, and Xiaomi.
The core breakthrough of WALL-B lies in its revolutionary shift from the VLA architecture to a unified world model architecture. Wang Hao, Co-founder and CTO of Independent Variable, pointed out at the launch that in traditional VLA models, the visual, language, and action modules are independent. Data loses information each time it is transferred between modules, causing the model to only "imitate" rather than truly "understand" the real world. WALL-B jointly trains all capabilities—vision, language, action, and physical prediction—from scratch within the same network, eliminating module boundaries and data transfer losses. Based on this architecture, WALL-B possesses three core features: vision, language, touch, and action are inherently integrated, enabling it to understand the handle position, opening direction, material, and even water content of a cup; it internalizes physical commonsense, capable of perceiving and predicting fundamental physical laws such as gravity, inertia, and friction; after task failure, it autonomously adjusts strategies, with successful experiences directly updating model parameters to achieve real-time self-evolution.
Wang Hao, Co-founder and CTO of Independent Variable, compared the clean, controllable data collected in laboratories to "sugar water data"—"It's sweet, but besides being sweet, it doesn't build resistance." Models trained on such data quickly fail in real-world environments. The noisy, variable data collected from real household environments are the "milk data" for embodied intelligence. To obtain this type of data, the Independent Variable team has entered hundreds of volunteers' real homes for model training. Wang Qian, Founder and CEO of Independent Variable, emphasized that the competitive barrier lies not in algorithm architecture or hardware, but in the complete engineering system encompassing data definition, collection, transformation, post-processing, and training evaluation adjustments. "In the robotics field, the leading window could be over three years."
Independent Variable has provided three clear solutions to address privacy concerns in household scenarios. The robot performs real-time pixelation of raw images on the device; raw images do not leave the device, achieving visual desensitization. The robot can only be activated after the user actively presses the consent button, eliminating "default consent" and ensuring transparent authorization. The robot never shares third-party data, recognizes only one owner, locks immediately upon detecting suspicious instructions, and has strictly defined purposes. Previously, Independent Variable collaborated with 58.com, deploying robots equipped with the WALL-AS model into real households to work alongside cleaning staff, verifying the feasibility of household scenarios.
Wang Qian also noted that the current model is still in the "intern" stage—it makes mistakes, requires remote assistance, and might sometimes put slippers in the kitchen or stop halfway through wiping a table to "think." However, it can work 24/7 without interruption and becomes "smarter" with each day of work due to the generation of new data. At the launch event, Independent Variable announced the immediate recruitment of "parents" for the first batch of household robots, with users able to submit applications through official channels. Wang Qian holds bachelor's and master's degrees from Tsinghua University. During his Ph.D., he conducted robotics learning research at a top U.S. robotics lab and was one of the earliest researchers globally to propose the attention mechanism in neural networks. Wang Hao holds a Ph.D. in computational physics from Peking University, previously served as the algorithm lead for the Fengshen Bang large model team at IDEA Research Institute, and led the development of China's first 10-billion-parameter large model, Ziya.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
