Vidu maker ShengShu releases Motubrain robot model

ShengShu Technology announced Motubrain, which it calls a general-purpose World Action Model built to serve as a single "robotic brain" in place of the task-specific or specialized systems that robots usually rely on.

The company says Motubrain ranks "highly" on WorldArena and RoboTwin 2.0, two benchmarks for embodied world models. The announcement does not give scores, rankings, or comparisons with named models, and it does not claim state-of-the-art results.

The release includes two scaling experiments. In the task-scaling test, each model was trained on data from a set number of tasks until its average success rate leveled off. ShengShu says Motubrain's success rate rose steadily as more tasks were added and that it scaled better than "conventional VLA baselines," which it does not identify. In the data-scaling test, demonstrations were subsampled evenly across tasks. The full dataset of 27,500 trajectories was trained for 50,000 optimization steps, and the company says performance kept improving as more data was added. The absolute success rates are not given in the text.

ShengShu presents Motubrain as an extension of its work on generative video. Its pitch is that video-based simulation lets robots learn from large and varied pre-training data and depend less on collecting data with physical robots.

"We believe general world models should not be built as stitched-together modules, but as a unified architecture that brings together perception, reasoning, prediction, generation, and action in a single system," said founder Jun Zhu.

ShengShu has not said whether the model will be available, which robots it supports, or whether it has been tested on real hardware.