The training process of large models is increasingly appearing in the public eye.
Previously, Luo Fuli, head of Xiaomi MiMo, made public the training progress of MiMo-V2.6's reinforcement learning stage, giving outsiders a glimpse into some specific details of large model training.
Now, a similar attempt has emerged in the Physical AI field.
On October 21, Shengshu Technology's world model team officially launched a public livestream of the entire Motubrain 2.0 training process, planned to last half a month, showing the public the training process and research progress of embodied models.
This livestream focuses on one question: how exactly is Physical AI achieved?
Physical AI, i.e., physical artificial intelligence, aims to enable AI to understand physical environments and complete real tasks through robots. Compared with large language models, Physical AI also needs to solve problems such as action decision-making, task execution, and cross-scene generalization.
The Motubrain series is a family of embodied models developed by Shengshu Technology, dedicated to building a general-purpose "brain" for robots and exploring capability generalization across different tasks, scenes, and robot embodiments.
Motubrain 2.0 is about to be open-sourced. Before the official open-source release, the team chose to first make public its entire training process, showing the outside world the model's journey from training and testing to capability iteration.
This public training covers four stages:
1. Pre-training: observing the impact of training scale on model performance and zero-shot generalization capability.
2. Post-training: further training on specific tasks to improve task execution capability.
3. Agent task execution: exploring 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.
What makes this livestream special is the attempt to open the "black box" of embodied model training.
In the past, the outside world learned about progress in embodied intelligence technology mainly through robot demos, evaluation results, and papers. What training a model went through and how its capabilities changed at different stages are usually hard to observe directly.
Motubrain 2.0 has chosen to make public the whole process from training and testing to capability iteration, hoping to let the outside world see how the model gradually forms capabilities, and also see the failures, trial and error, and adjustments during training.
For the embodied intelligence industry, this approach makes the model development process more transparent. Researchers and developers can not only see the final results but also have the opportunity to understand the specific problems encountered in training and how the team continuously improved the model through experiments and feedback.
At the same time, the team also hopes to use this livestream to share research experience, problems, and reflections with technical peers, jointly discuss key issues such as large-scale training, cross-task generalization, and autonomous improvement, and provide more references for exploring paradigms for realizing Physical AI.
Of course, the related research is still at an exploratory stage, and the specific results still need to be verified by follow-up experiments and evaluations. Compared with only showing what a robot can ultimately do, making public how the model is trained and how its capabilities form is equally valuable for research.
Livestream entrance: www.motubrain.com/motubrain2.0-live
Official website link: www.motubrain.com