LEGO: Latent-space Exploration for Geometry-aware Optimization of Humanoid Kinematic Design

A research team from Korean institutions including Jihwan Yoon, Sungjoon Choi, and six others has introduced LEGO (Latent-space Exploration for Geometry-aware Optimization), a novel AI framework that fundamentally changes how humanoid robots are designed. Published on arXiv and accepted for ICRA 2026, the system addresses two major bottlenecks in robot design: the vast, unstructured nature of the design space and the difficulty of creating task-specific loss functions. Instead of relying on human intuition or hand-crafted parameters, LEGO learns a compact search space directly from existing mechanical designs using isometric manifold learning and screw-theory-based joint axis representation. This creates a geometry-preserving latent space where optimization becomes tractable.

The framework then defines its optimization objective directly from human motion data through motion retargeting and Procrustes analysis, essentially teaching robots to move by mimicking human biomechanics. Using gradient-free optimization within this learned latent space, LEGO can automatically discover novel robot kinematics tailored to specific tasks. This represents a shift toward fully data-driven robot design, where the AI system leverages both existing engineering knowledge (encoded in prior designs) and biological inspiration (from human motion capture) to generate optimized robotic morphologies that would be difficult for human designers to conceptualize systematically.