Lumo-1

Lumo-1 is Astribot's first generalist vision-language-action model, published with a technical report on 9 December 2025. Its defining trait is that reasoning is explicit: the model produces a written chain of thought about the scene and the task, and only then emits the motion, which is what the company means by connecting "mind to hand". The backbone is Alibaba's Qwen2.5-VL-7B, a 7-billion-parameter vision-language model. Training runs in three supervised stages on 128 H100 GPUs — 13.7 billion tokens of curated vision-language data, then 200 billion tokens of cross-embodiment robot data covering 145 tasks on machines including AgiBot Genie-1, an Astribot S1 prototype and bimanual ARX, YAM and AgileX arms, then 193 billion tokens of action training on 16.2 million teleoperation frames recorded on the S1 itself. A reinforcement-learning stage with GRPO closes the loop between what the model says it will do and what it does. The technical contribution the company highlights is its spatial action tokeniser, which compresses a motion into the shortest sequence of waypoints instead of binning joint values, so the same representation carries across different robot bodies. On embodied reasoning benchmarks the report puts Lumo-1 at 86.36 on CV-Bench and 87.90 on BLINK, ahead of RoboBrain-7B-2.0 in most of the suite. The weights are not published; the model is documented in a research report and demonstration videos, and runs on Astribot's own S1 manipulator. It was superseded in July 2026 by Lumo-2, which drops the written plan in favour of prediction in latent space.

▸Videos

No videos yet.

▸Reviews