X Square Robot Open Sources Wall-OSS-0.5 for Embodied AI Research

X Square Robot announced in a press release the open-source release of Wall-OSS-0.5, a Vision Language Action model built for real-world robotic manipulation. The release aims to show that pretrained robotic models can perform directly on physical robots without additional fine-tuning.

Wall-OSS-0.5 was evaluated on a 17-task real-robot zero-shot suite covering semantic understanding, object manipulation, and long-horizon control. The pretrained model achieved task progress scores above 80 on several tasks, including Block Sorting, Fruit Sorting, and Rope Tightening. These results suggest measurable robot behavior from pretraining alone.

The model introduces gradient-bridged co-training, which integrates robotic action supervision into the vision-language backbone. It also incorporates a Vision-Aligned RVQ Action Tokenizer and Action-Space Supervision to improve action representation and training stability. For large-scale optimization, X Square Robot developed DMuon, a distributed optimizer that reduces computational overhead.

The company is releasing the full Wall-OSS-0.5 stack, including model weights, training code, and recipes, to support reproducible research in embodied intelligence and enable direct testing of pretrained robotic behaviors in real-world settings.