X Square Robot Launches WALL-B For Home Deployment

X Square Robot announced `WALL-B`, an embodied AI foundation model built on a bold `World Unified Model (WUM)` architecture, and said a new robot generation using the model and upgraded hardware will begin home deployments in 35 days. The company highlights joint training across perception, language, motion, and physics and says training data includes experimental scenarios plus recordings from hundreds of households. Privacy measures cited include on-device visual anonymization, explicit user authorization, and usage limits. Further technical details will be disclosed on April 27 at the Guangdong AI Application Conference.

The headline technical pivot is moving away from modular Vision-Language-Action pipelines to a natively fused multimodal foundation model. `WALL-B` is described as jointly optimizing several capabilities, which X Square frames as overcoming information loss between modules and enabling direct physical reasoning. Core capabilities the company lists are:

• •native multimodality, combining vision, language, motion, and prediction in a single training objective
• •physical world dynamics modeling, enabling anticipatory motion and manipulation planning
• •self-evolution after failure, implying online adaptation or improved recovery policies after execution errors

Training reportedly blends controlled experiments and in-home data to close the sim-to-real gap. The announcement also pairs the model release with hardware upgrades for perception and manipulation, though public details on sensors, compute, or on-board inference stacks are not yet available. Privacy and safety controls are emphasized, with on-device visual anonymization called out as the primary mitigation.

This is a practical push to move embodied AI from lab demos into consumer settings. The WUM framing echoes a broader trend toward unified state representations and end-to-end learning for perception-to-action tasks. If WALL-B genuinely learns cross-modal physical dynamics from real household distributions, it could shorten adaptation time for diverse home layouts and tasks. However, claims of deployment in 35 days are aggressive; real-world robotics trials typically surface edge cases in perception, manipulation, and human interaction that need extended validation. Privacy, safety, and long-tail robustness will determine whether pilots scale.

Inspect the April 27 technical release for architecture diagrams, dataset composition, training regimes, evaluation benchmarks, hardware specs, and third-party validation. Monitor early pilot reports for failure modes, latency of on-device inference, and the effectiveness of the stated privacy controls.