Behind the Robots Assembling the Great Wall

机器之心 · wechat · 2026-07-17

Force Lingji and StepStar showcased an impressive robot assembly project ahead of WAIC: **6 台机器人连续工作 15 小时** to build a Great Wall model comprising **8 万多个零件**. The smallest component was less than 1 centimeter, and the final model measured 3.5 meters long, 1.5 meters wide, and 1.1 meters high. The division of labor involved 4 desktop robots for precise assembly and 2 wheeled robots for transport and assembly. The demonstration wasn't just for show, but highlighted system integration: **DM0.5** served as the general foundation model, paired with the **DW0.5** world model for a reinforcement learning closed loop. The article notes DM0.5 excels in long-horizon tasks, memory fusion, instruction understanding, and dynamic trajectory alignment. Meanwhile, DW0.5 acts as a "Learned Environment" during post-training, simulating both success and failure trajectories to provide denser value feedback. In evaluations, DM0.5 topped multiple robotics and navigation benchmarks, while DW0.5 achieved SOTA in world model and action policy metrics. The core takeaway is that the next step for embodied AI isn't just scaling parameters, but combining foundation models, world models, and post-training loops to deliver tangible real-world results.

Related event: 6 Robots Autonomously Build 80,000-Block Great Wall in 15 Hours(2 posts)→

Original post →

More from Companies & People

Companies & People channel →