WALL-A

WALL-A is an embodied “operation model” (操作大模型) developed by X Square Robot (自变量机器人, Shenzhen), serving as the core of the WALL family of foundation models. The company describes it as one of China’s earliest fully end-to-end implementations of embodied AI, integrating perception, understanding and action control in a single model.

The model uses an end-to-end architecture at a tens-of-billions parameter scale (百亿级, per the vendor’s materials) and leverages the World Unified Model architecture together with the X-Tokenizer action tokenizer. It takes vision (RGB-D, 3D-ToF sensors), the robot’s proprioceptive data and natural-language instructions as input, and produces action-control sequences as output.

WALL-A supports few-shot learning across tasks, manipulation of difficult objects (friction, fluids, deformable objects), force control and precise positioning, embodied chain-of-thought reasoning, and long operation sequences with sub-millimeter precision on high-degree-of-freedom systems.

The model is deployed in X Square Robot’s robots, including the Quantum X1 Pro (wheeled dual-arm), the Quantum X2 (wheeled humanoid) and the five-finger ArtiXon Hand. It is a proprietary (closed) solution available through the company’s products rather than as a separately downloadable model.