Shengshu Opens a Two-Week Public Livestream of Motubrain 2.0 Embodied-Model Training

Shengshu Technology's world-model team began a public livestream of the full Motubrain 2.0 training process on October 21, according to a QbitAI report, with the broadcast planned to run for half a month.

Motubrain is Shengshu's embodied-model line, built to give robots a general-purpose brain and to test how well capability transfers across tasks, scenes and robot bodies. Motubrain 2.0 is due to be open-sourced, and the team chose to open the training process before the release rather than after it.

The stream covers four stages: pretraining, where the team observes how training scale affects model performance and zero-shot generalization; post-training on specific tasks to improve execution; agent task execution, testing whether learned experience transfers to other tasks; and an RSI, or recursive self-improvement, exploration of whether the model can evaluate and improve its own actions from execution feedback.

What makes the broadcast unusual, the report argues, is the attempt to open the black box of embodied-model training. Progress in embodied AI has generally been communicated through robot demos, benchmark results and papers, with little visibility into how a model was trained or how its abilities changed between stages.

Publishing the path from training through testing to capability iteration means viewers can see abilities form, including the failures, retries and adjustments along the way. For researchers and developers, those failure cases are often more useful than a polished demo.

There is precedent. The report notes that Luo Fuli, who leads Xiaomi's MiMo effort, previously disclosed training progress from the reinforcement-learning stage of MiMo-V2.6, giving outsiders a look at concrete details of large-model training, while physical AI had no comparable attempt.

The report also carries the necessary caveat: the work remains exploratory and results still need validation through later experiments and evaluations. The team frames its focus as how the model trains and how capability forms rather than the final demo.

QbitAI notes at the end of the piece that the article was supplied by Shengshu Technology and republished with permission, so independent observation of the stream still matters. For the embodied AI field, opening the training process reduces information asymmetry and puts scaling, cross-task generalization and self-improvement up for public discussion. What to track is how the four stages actually progress over the two weeks, and under what weights and license Motubrain 2.0 ships.