Astribot Launches Lumo-2: A World-Action Model on Frozen Qwen VLM, Robot Learns Latte Art with Thin Adapters

Chinese robotics company Astribot released Lumo-2, a World-Action Model based on a frozen Qwen-3.5 4B VLM, trained in three progressive stages: aligning actions with latent world dynamics, aligning with vision-language, and end-to-end training. While not top on benchmarks, its novelty lies in combining latent world modeling with action generation and leveraging unannotated human video for generalization. With only thin trainable adapters, it's suitable for real-time embedded inference and excels at long-horizon tasks.

• Qwen
• world model
• Embodied AI
• robot
• action model
• Astribot
• Lumo-2