X Square Robot Proposes Integrated Stack for General-Purpose Robots
Chinese embodied AI company X Square Robot is advocating for an integrated approach to building general-purpose robots, contrasting with traditional methods that assemble separate perception, planning, and control modules. The company proposes a "stack" encompassing robot learning data, a world model for predicting physical changes, and an action model that unifies perception, planning, reasoning, and decision-making. This stack is guided by principles emphasizing interactions as the basic unit of robot data, pretraining for usable capability, and behavior modeling around physical events rather than fixed time slices.
X Square Robot has developed the Universal Manipulation Interface (UMI) data collection system, QUANXTA Zero Series, to address the cost and quality of interaction data. This system uses a wearable rig for human demonstrations and incorporates a unique quality control loop with physical playback on a real robot to validate successful task completion. By combining large volumes of robot-free demonstrations with a small amount of real-robot data, the company claims to achieve comparable performance to all-robot datasets at significantly lower costs. Their world model, WALL-WM, is designed around action-grounded semantic events, enabling variable-length reasoning for long horizons and fixed-length output for real-time control, preserving visual priors from text-to-video models.
The action layer of their proposed stack, embodied in Wall-OSS-0.5, aims for a pretrained model that can operate on a real robot without task-specific fine-tuning. This model trains discrete action tokens, language grounding, and continuous action generation simultaneously. The associated action interface, X-Tokenizer, creates semantically meaningful action tokens that align with language model features, enhancing stability and interpretability. While promising, the company notes that strongest results are currently measured on their own hardware and pipelines, with broader independent testing needed to confirm wider applicability.
The proposed integrated stack by X Square Robot represents a significant departure from modular robotics architectures, aiming to imbue robots with more generalizable intelligence akin to large language models. By prioritizing interaction quality over sheer data volume and focusing on event-grounded world models, the company addresses key limitations in current embodied AI development. The emphasis on open-sourcing their stack suggests a strategy to foster community-driven advancement and validation, potentially accelerating progress in the field. Future challenges will involve demonstrating the stack's efficacy across diverse hardware platforms and real-world scenarios beyond their controlled environments, as well as navigating the complexities of integrating diverse data modalities for robust, adaptable robotic agents in the coming decade.
AI-generated to prompt reflection — not editorial opinion, not advice, not a statement of fact. How this works.