Theo Jansen's Strandbeest (Beach Monster) is a wind-powered walking mechanical device that converts rotational motion into biomimetic leg-like gaits through intricate linkage mechanisms. Now, developers have attempted to build a 3D simulator of this complex mechanism using AI.


Developer Dilum Sanjaya used Gemini 3 to build the core simulator and Nano Banana to generate the user interface. The test goal was to verify whether AI could first construct a single-legged linkage mechanism and then correctly expand it into a full multi-legged walker.

The experimental results showed that both Claude Opus 4.5 and Gemini Pro failed when instructed to "create a 3D Strandbeest simulator with 5 legs." Even after multiple prompts, the models could not directly complete this complex task.

The developers then simplified the requirement, asking AI to build a 3D model of a single-legged mechanism. After several attempts, both models successfully generated functional single-legged mechanisms. However, the real challenge lies in scaling: replicating and coordinating single-legged mechanisms into a multi-legged system.

Gemini eventually completed the expansion after several iterations, while Claude Opus 4.5 required manual code review to identify specific errors and continue. Even so, both models failed to solve a critical issue: certain linkages collided with the central beam. In actual Strandbeest designs, clever geometric structures are used to avoid such interference.


This experiment is the sixth part of the "Vibe Coding Robotics" series. In previous works, the developer demonstrated how to redesign the interface of a robot simulator using Nano Banana. By capturing screenshots of old applications and letting AI add relevant widgets, multiple UI variants could be quickly generated, and the most suitable design could be selected.



The developer mentioned that choosing Nano Banana over Gemini 3 to generate the UI had two main reasons: faster speed, allowing multiple UI variants to be produced within the time it took Gemini to generate once; and each time Gemini was asked to regenerate a widget, it tended to completely replace existing content rather than incrementally improve it.

This case illustrates the current capabilities and limitations of AI in handling complex engineering problems. AI can perform relatively independent tasks well, but in complex projects requiring systems thinking and multi-step coordination, human intervention and guidance are still necessary. For tasks like mechanical design and motion coordination, which require precise geometric calculations, AI's performance is more like that of a junior engineer who needs strict supervision.

The final generated Strandbeest simulator, while functionally complete, still falls short of perfection. This reminds us that when applying AI to engineering fields, we must have a clear understanding of its capabilities—it can accelerate the development process but cannot fully replace human professional judgment.