4B VLA Model 'Wall-OSS-0.5' Released

Wall-OSS-0.5 is a 4B-scale VLA (Vision-Language-Action) model that combines a 3B VLM backbone with action experts in a Mixture-of-Transformers structure.

Key performance metrics are as follows:

• Zero-shot performance: Evaluated on a set of 17 real-robot tasks, achieving a progress score of 80 or higher on 4 tasks, including deformable tasks such as 'Rope Tightening'.
• Fine-tuning performance: Achieved an average progress score of 60.5 when fine-tuned on a set of 15 tasks, a 17.5pp improvement over pi0.5.
• Embodied grounding: Improved related performance by 21.8pp while maintaining general VLM capabilities.

Key technical elements include the Gradient Bridge concept, in which the CE (Cross-Entropy) of action tokens provides the dominant gradient to the VLM backbone, and the Vision-Aligned RVQ Tokenizer, which semantically grounds action tokens. It also introduces a decentralized DMuon optimizer with reduced overhead.