Flow - 4x faster than typing
Mini toolsLumo-2
By Astribot
Lumo-2 is a latent world-action model built through a three-stage progressive alignment process that links action representations to visual world dynamics, then to vision and language semantics, then to joint training across vision-language, video, and robot data. It uses block-wise autoregressive decoding for real-time closed-loop control and a lightweight temporal memory to handle long-horizon, multi-phase tasks. Astribot reports it outperforms baselines such as Pi-0.5 and Fast-WAM on temporal reasoning, physical understanding, and dexterous manipulation benchmarks, and supports cross-embodiment transfer from human video data without specialized transfer learning methods.
New
Multimodal
Overview
Lumo-2 is Astribot's second-generation latent world-action model for humanoid robots. It predicts short-term, action-relevant physical changes in a compact latent space before generating motor actions, replacing explicit text-based planning with implicit predictive reasoning grounded in visual observation, robot state, and language instructions.
Last updated: July 15, 2026
