Shengshu Technology livestreams full Motubrain 2.0 Physical AI training for about half a month

Shengshu Technology (生数科技) began a public livestream of the full training process for Motubrain 2.0 on October 21, according to a post by the company that Chinese outlet QbitAI (量子位) republished with authorization. The stream is planned to last about half a month. Motubrain 2.0 is an embodied model for robots, and the company says it will be open-sourced soon. The source article is in Chinese.

What is being streamed

The livestream is run by Shengshu's world model team and is meant to show how an embodied model is trained and how its research progresses. The company describes four stages:

1. Pretraining: observing how training scale affects model performance and zero-shot generalization.
2. Post-training: further training on specific tasks to improve task execution.
3. Agent task execution: testing whether the model can transfer existing experience to other tasks.
4. RSI (recursive self-improvement) exploration: studying whether the model can evaluate and improve its own actions based on execution feedback.

The team says it will show the process from training and testing through to capability iteration, including failures, trial and error, and adjustments. The livestream is available at www.motubrain.com/motubrain2.0-live. The company's main site is www.motubrain.com.

About Motubrain

The Motubrain series is an embodied model line developed by Shengshu. The company says it aims to build a general-purpose "brain" for robots and to explore how capabilities generalize across tasks, scenarios and robot bodies.

The company defines Physical AI as AI that understands physical environments and completes real tasks through robots. Compared with large language models, it says, Physical AI must also handle action decision-making, task execution and cross-scenario generalization.

Shengshu says the stream is also meant to let the team share experience, problems and thinking with peers on scaling training, cross-task generalization and autonomous improvement.

Context

QbitAI notes that public views of large model training are becoming more common. It cites Luo Fuli (罗福莉), who leads Xiaomi's MiMo effort and earlier shared training progress from the reinforcement learning stage of MiMo-V2.6. QbitAI describes the Motubrain livestream as a similar attempt in the Physical AI field.

QbitAI says progress in embodied intelligence has mostly been visible through robot demos, evaluation results and papers, and that it is usually hard to observe how a model was trained or how its capabilities changed across stages. Shengshu's stated aim is to open up this "black box."

The company notes the research is still exploratory, and that specific results will need to be confirmed by later experiments and evaluations.

What this means

The livestream shifts the visible evidence for a Physical AI model from finished demos to the training process itself. If the stream covers all four stages as described, observers will be able to see how scale affects zero-shot generalization and how post-training changes task performance. They will also see whether experience carries over to new tasks. Shengshu says the stream will include failures and adjustments, which are rarely visible in demos, evaluation results or papers.

The RSI stage is the most exploratory item on the list, and the company itself says outcomes still need validation. No benchmark scores, parameter counts or open-source release date for Motubrain 2.0 have been stated, so the model's capabilities can only be judged from what the livestream shows and from later evaluations. The planned open-source release would then let researchers test the model directly.