Shengshu Technology Opens Half-Month Motubrain 2.0 Training Livestream
Shengshu Technology began a half-month livestream of Motubrain 2.0's full training before open-sourcing the embodied model.
Physical AI aims to let AI understand physical environments and perform real tasks through robots. Compared with large language models, it must also address motion decisions, task execution, and cross-scenario generalization. Motubrain is Shengshu's embodied model series intended to give robots a general-purpose "brain" and test whether capabilities can generalize across tasks, scenes, and robot bodies.
Before the official open-source release, the team is making the model's training, testing, and capability-iteration process public. The livestream covers pretraining, to observe how training scale affects model performance and zero-shot generalization; post-training on specific tasks to improve execution; agent task execution, to explore whether the model can transfer existing experience to other tasks; and RSI (recursive self-improvement) exploration, to study whether the model can evaluate and improve its actions based on execution feedback.
The team says the livestream is meant to open the "black box" of embodied-model training. Previously, outsiders mainly learned about embodied AI through robot demos, benchmark results, and papers, while the training process and capability changes at different stages were hard to observe directly. By showing training, testing, and iteration, the team says it wants viewers to see how capabilities form, including failures, trial and error, and adjustments.
The approach could make model research and development more transparent for researchers and developers, who can see final results as well as specific problems and how the team used experiments and feedback to improve the model. Shengshu also says it hopes to share experience and questions with peers on scaling training, cross-task generalization, and autonomous improvement, and to provide references for exploring Physical AI paradigms. The company notes the research remains exploratory and that results still need to be verified by later experiments and evaluations.
The livestream follows a similar move in large-model training: Leiphone reported that Luo Fuli, who leads Xiaomi's MiMo, publicly shared training progress in the reinforcement-learning stage of MiMo-V2.6, giving outsiders a view of specific training details. Shengshu is now trying a comparable public process in Physical AI. The company's website is www.motubrain.com.