X Square Robot Open Sources WALL-WM, an Event-Based World Model for Robots

X Square Robot announced in a press release the open-source release of WALL-WM, a world action model designed for general purpose embodied AI. The model focuses on teaching robots to understand and learn from discrete physical events rather than fixed time slices.

WALL-WM organizes both training data and supervision around action-grounded events such as reaching, grasping, and placing. Each event is represented through language, video, and action data, forming a unified structure for robot learning. This approach replaces the conventional method of training on equal-length time chunks, allowing robots to learn based on meaningful changes in the environment.

The system uses a prior-aligned video-action architecture that connects a pretrained video model with a newly initialized action module. It supports multi-view perception through geometry-aware training and cross-view attention, enabling robots to handle multiple camera inputs without additional calibration.

WALL-WM can operate in two modes. Event Mode executes variable-length event segments guided by language or visual cues, while Unified Mode retains fixed-length inference conditioned by event reasoning. The model’s training pipeline integrates diverse data sources, including internet videos, teleoperation data, and robot datasets, annotated across different temporal scales.

Built with distributed Muon optimization and FP8 quantization, WALL-WM improves computational efficiency during large-scale training and inference. Benchmark results show stronger physical prediction and generalization across scenes and tasks compared to previous models. The open-source release includes code and resources on the company’s GitHub repository.