X Square Robot builds foundational stack for general-purpose robots

Chinese company X Square Robot, specializing in embodied AI, has proposed an integrated stack for building general-purpose robots. Unlike the traditional approach where systems are assembled from separate perception, planning, and control modules, X Square Robot integrates data, a world model, and an action model into a single open platform. The stack is based on three principles: the basic unit of data is interaction, not trajectory; pretraining should provide usable ability, not just initialization; behavior should be modeled around physical events, not fixed time segments. For data collection, the company developed the Universal Manipulation Interface (UMI) QUANXTA Zero Series, which uses a wearable helmet with two grippers instead of teleoperation. The key innovation is quality control: recorded trajectories are replayed on a real robot, and only those that successfully complete the task are considered valid. The world model WALL-WM operates on semantic events (e.g., "grasp", "place") rather than fixed time segments and uses a text-to-video model while preserving visual features. The action model Wall-OSS-0.5 is trained on three tasks simultaneously: discrete action tokens, language grounding, and continuous action generation. The tokenizer X-Tokenizer converts continuous motions into semantic tokens, where the higher-order code represents motion intention and lower-order codes represent details, ensuring transferability between robots. The company reports results on its own benchmarks, and the world model code is already being released for independent testing. X Square Robot's valuation has exceeded 20 billion yuan (about 2.9 billion US dollars), indicating investor confidence in the company's approach.