While human engineers have spent decades crafting robots with straight edges and rectangular parts, artificial intelligence has just blown their conventional designs to smithereens. Diffusion models are now generating robot shapes that look nothing like traditional designs. Think curved linkages instead of straight ones. Blob-like appendages rather than neat rectangles. The nerds are shook.
These AI systems sample 500 potential designs through embedding vectors, narrowing down to the top 12 for optimization. The process repeats five times, each iteration making the robots jump higher. The final designs? They look like thick drumsticks or amorphous blobs. Not exactly what you’d sketch on your engineering pad.
Nerds trembling as AI spits out blob-shaped robots that utterly demolish traditional engineering assumptions
The results speak for themselves. These weird-looking robots leap 41% higher than conventional designs. Forty-one percent! That’s not a small improvement—it’s a transformation. The curved, blob-like structures store energy more efficiently, creating robots that jump higher and land more safely. The improved foot design resulted in an 84% improvement in stability when landing after jumps. The AI deliberately avoids thin, lightweight links that might snap under pressure. Smart move.
Unlike RFM-1, which uses 8 billion parameters to enhance reasoning capabilities, these systems prioritize physical form over computational complexity. What’s fascinating is how the AI approaches problems differently. Similar to how computer vision provides instant feedback on athlete techniques, these AI systems continuously refine robot designs through visual analysis. No human would start by saying, “Let’s make robot legs that look like drumsticks.” But AI doesn’t care about convention. It only cares about physics and performance. And it works.
Training happens fast, too. Hours instead of weeks. The robots learn through a combination of human demonstrations and self-experience, mastering complex tasks with 100% success rates in some cases. They’re assembling motherboards, building shelves, and even doing parkour across difficult terrain.
These machines reason about their environment in near-human ways, handling deformable objects and adapting to varying tasks on the fly. The integration of spatial reasoning and coding abilities allows robots to instantiate new skills dynamically.
Engineers can only watch in awe as their straight-edged assumptions crumble. Sometimes, you need AI to tell you your decades of design principles were just… wrong.
References
- https://news.mit.edu/2025/using-generative-ai-help-robots-jump-higher-land-safely-0627
- https://covariant.ai/insights/introducing-rfm-1-giving-robots-human-like-reasoning-capabilities/
- https://news.berkeley.edu/2025/01/28/using-ai-these-robots-learn-complicated-skills-with-startling-accuracy/
- https://deepmind.google/discover/blog/gemini-robotics-brings-ai-into-the-physical-world/
- https://www.rudebaguette.com/en/2025/06/robot-defies-physics-new-agile-quadruped-parkours-over-gaps-and-walls-with-terrifying-precision-that-stuns-ai-researchers/