Wall-OSS

Wall-OSS is an open-source end-to-end embodied foundation model released in September 2025 by Independent Variable Robotics (自变量机器人), a portfolio company of Yunqi Capital . It represents a lightweight, open version of the company's proprietary WALL-A model, which was built on a VLA (vision-language-action) architecture .

The model was designed with a "shared attention + expert shunting (FFN)" architecture that transfers VLM knowledge losslessly into manipulation models, and a three-stage training paradigm of "first discrete, then continuous, then joint" training . Beyond conventional robot control, WALL-OSS incorporates long-sequence chain-of-thought reasoning, language and sub-task capabilities, and was positioned by the company as having leading zero-shot CoT performance comparable to Physical Intelligence's models .

The release came amid a broader push into open-source embodied intelligence in 2025, with Physical Intelligence open-sourcing its own model one day later . Independent Variable Robotics has since evolved its architecture approach, moving from VLA toward a "World Unified Model" (WUM) architecture with its subsequent WALL-B and WALL-WM releases in 2026 .