Shengshu’s Motubrain Team Opens Two-Week Live Stream of Physical AI Training

The stream is focused on Physical AI, or physical artificial intelligence, which aims to let AI understand physical environments and complete real tasks through robots. Compared with large language models, the report says Physical AI also has to address action decisions, task execution and cross-scene generalization.

Motubrain is Shengshu’s embodied model series, intended to build a general-purpose “brain” for robots and explore how capabilities can generalize across tasks, scenes and robot bodies. Motubrain 2.0 is due to be open-sourced. Before that release, the team chose to make its full training process public, from training and testing to capability iteration.

The public training covers four stages. In pretraining, the team will observe how training scale affects model performance and zero-shot generalization. In post-training, the model is trained further for specific tasks to improve task execution. In agent task execution, the team will explore whether the model can transfer existing experience to other tasks. In an RSI, or recursive self-improvement, exploration, it will study whether the model can evaluate and improve its own actions based on execution feedback.

QbitAI reported that the live stream is an attempt to open the “black box” of embodied model training. Previously, outsiders mainly learned about progress in embodied intelligence through robot demos, evaluation results and papers; the training a model underwent and how its abilities changed at different stages were usually difficult to observe directly. Motubrain 2.0’s public process is meant to show how abilities form step by step, as well as failures, trial and error, and adjustments during training.

For the embodied intelligence industry, the company says the approach makes model research and development more transparent. Researchers and developers can see not only final results but also specific problems encountered in training and how the team used experiments and feedback to improve the model.

The team also says it wants to use the stream to share experience, problems and thinking with technical peers, and to discuss large-scale training, cross-task generalization and autonomous improvement, offering references for exploring a paradigm for Physical AI. The research remains exploratory, and its effects still need to be verified by later experiments and evaluations. The company says that compared with showing only what a robot can ultimately do, disclosing how a model is trained and how capabilities form also has research value.

The report was provided by Shengshu Technology and republished by QbitAI with authorization, according to the article.