The Chinese company ShengShu Technology has introduced a new artificial intelligence (AI) system called Motubrain, a universal model that the developers call a single brain for robots, Interesting Engineering reports.

Unlike traditional robotic systems, where perception, planning, and actions are divided between different software modules, Motubrain combines all this within a single architecture.

The project authors claim that the model is simultaneously trained on the basis of videos, text instructions, and physical actions. Thanks to this, the robot can not only recognize the environment, but also predict the development of the situation, make decisions, and immediately perform actions without switching between separate systems.

Motubrain is based on an architecture that the developers call Mixture-of-Transformers. It combines image, language, and motion processing in a single cycle. According to Jun Zhu, Founder of ShengShu Technology, the goal of the project is to create a system that is not a set of stitched modules, but a complete model of the real world, capable of understanding cause-and-effect relationships and the physics of the environment.

One of the main features of this system has become training on a huge amount of unlabeled video. Instead of manually marking millions of actions, the model independently extracts movement patterns from human videos, simulations, and the actions of other robots. This significantly speeds up the scaling of training.

The developers claim that Motubrain can perform chains of up to ten consecutive actions, while many modern robotic systems confidently overcome only two or three steps in a row. During the tests, robots with the new AI were able to adapt to unexpected situations: for example, if the robot unsuccessfully grabbed an object, the system independently recognized the error and repeated the experiment, although it was not specifically trained for this scenario.

The model has shown high results on test platforms. In the WorldArena benchmark, it scored 63.77 points, and in RoboTwin 2.0, it received an average of 96 points out of 100 for 50 different tasks. The developers claim that this is the only system that has overcome the 95-point threshold in randomly changing environmental conditions.

Motubrain is now already being used in robot training programs for industrial, commercial, and household tasks. ShengShu Technology is collaborating with several robotics companies, and the development is seen as a step towards the creation of universal “embodied AI” systems that will be able to operate in the physical world almost as flexibly as a person.