Ola-Sem and HKUST(GZ) Open-Source Embodied Model UniWAM, Revealing Scaling Trends in Human-Robot Co-training

Euler Everything and The Hong Kong University of Science and Technology (Guangzhou) have jointly released UniWAM (Unified World and Action Model), the open-source version of Ola-Sem; the model and training code, along with pretrained and post-trained weights, have all been open-sourced. The question it attempts to answer is: can a robot stop treating understanding, prediction, and action as separate capabilities, and instead learn them within a unified architecture, thereby reducing the disconnect between understanding and execution.

UniWAM integrates three types of components into a unified architecture: a physical reasoner based on Qwen3-VL-2B-Instruct is responsible for understanding the current visual scene and task instructions, providing semantic information; a world generator based on Wan2.2-TI2V-5B learns to predict future visual changes, providing dynamic information for manipulation; and an action predictor models the flow field of continuous actions, generating the action segments needed for robot execution. The three exchange information through joint multimodal attention, enabling action prediction to combine task requirements with scene dynamics.

This design means UniWAM's capabilities are constrained by the foundation models it relies on. The physical reasoner inherits the semantic capabilities of Qwen3-VL-2B-Instruct, the world generator inherits the spatiotemporal priors of Wan2.2-TI2V-5B, and the action predictor learns continuous control signals on top of these. The three components retain their respective feature processing methods, then project tokens from different modalities into a shared attention space.

The team self-reports that UniWAM achieved an average success rate of 99.2% on the LIBERO benchmark, an overall success rate of 92.6% on LIBERO-Plus, which introduces various environmental perturbations, and a success rate of 68.32% on the RoboTwin 2.0 randomized environment test (C2R). The SOTA rankings for these three results are limited only to the highest values among the methods listed in the corresponding tables of the paper, and are not a comprehensive comparison against all published methods.

LIBERO-Plus evaluates model robustness across seven dimensions: background, camera viewpoint, language instructions, lighting, object layout, robot initial state, and sensor noise. The team self-reports that UniWAM achieved the highest scores among the methods listed in the paper in four evaluations—robot initial state, language, lighting, and background—and reached an 84.6% success rate in the object layout variation evaluation.

Switching from the standard environment (C2C) to the randomized environment (C2R), the team self-reports that UniWAM's success rate dropped by only 6.82 percentage points, and it achieved the highest score among the listed methods in the C2R test. This drop is used to support the model's generalization ability to environmental changes, but it is likewise limited by the comparison scope of the methods listed in the paper's tables.