Lumo-2

Lumo-2 is the second generation of Astribot's foundation model for robot manipulation, published with a technical report on 13 July 2026. Where the first generation wrote out its plan in text before moving, Lumo-2 reasons in a latent space: it learns how a scene is likely to change and picks actions against that prediction, without spelling the plan out in words. The backbone is Qwen3.5-4B, a 4-billion-parameter vision-language model from Alibaba, extended with a vector-quantised action codebook. Training runs in three stages — the company reports 30,000 steps on 64 H100 GPUs, then 12,000 on the same cluster, then 120,000 steps on 160 H100s. The vision-language mix comes to roughly 53 million samples; the robot data mixes Astribot's own recordings from its S1 manipulator with public cross-embodiment sets, including AgiBot Genie-1, Galaxea R1 Pro and tabletop arms from ARX and AgileX, plus egocentric human video from Ego4D and EPIC-KITCHENS. The headline practical number is latency: block-wise autoregression brings end-to-end inference down to 93.5 ms, which the authors put at 2.71 times faster than standard autoregressive decoding. On pick-and-place the report claims a 96.7 percent success rate in the basic setting and 85 percent on objects the model has never seen, against 70 percent for π0.5 in the same test. A 22-task real-world suite covers temporal reasoning, physical understanding and control complexity, with tasks such as making coffee and packing a suitcase. All of these figures come from Astribot's own evaluation and have not been reproduced independently. The weights are not published: the model is described in a research report and shown in videos on the company's site, and access is limited to Astribot's own robots.

▸Videos

No videos yet.

▸Reviews