A lot of work being coming out on robot foundation models.

Astribot just released Lumo-2, a 4B robot foundation model that learns to predict the real world before deciding how to move.

• 2.71x faster total inference,
• support for human videos and multiple robot bodies,
• stronger results on 105 unseen objects, and
• best overall performance across 22 real-world manipulation tasks covering motion prediction, memory, physical reasoning, long tasks, and fine hand control.

And these results were validated on a real robot across 20+ complex household tasks, not just benchmark leaderboards.

That is a much harder test because the model has to deal with changing scenes, physical contact, timing, memory, and long action sequences where one small mistake can break the entire task.

Lumo-2 is a robot control model with a small world model built inside it, I could say its industry's first Latent World-Action Model (LWAM) built for home.

In a normal world model, we predict future images or video, and then a separate planner decides what the robot should do. In a normal robot policy, we skip prediction and directly convert the camera view and instructions into motor commands.

So Lumo-2 does both those things.

First, it only predicts the part of the future that will affect the task, like an object moving, a contact changing or the pouring being finished. Second, it uses that small prediction to move the robot. It doesn’t create an entire video from the future, frame by frame.

The model is trained in three steps. First, it links changes of perception to robot motion. Second, it grounds action tokens in language and vision. Third, it is trained on VLM data, videos and robot demonstrations, such that the sequence can anticipate the physical outcome and then act.

🧵 1.