Neural operators generalize continuum robot designs for faster control

Continuum robots offer dexterous manipulation in tight spaces but require fast, accurate models for real-time control. Traditional physics-based models are computationally heavy and often inaccurate, while learning-based methods fail to generalize beyond the specific robot they were trained on. To solve this, Branden Frieden and colleagues re-framed the surrogate modeling task as an operator learning problem: mapping robot design parameters and tendon actuation inputs directly to resulting configurations. This approach lets a single trained model work across many different robot designs.

They built and tested four neural operator architectures—two based on Deep Operator Networks (DeepONets) and two based on Fourier Neural Operators (FNOs)—training them on simulation data. All four achieved strong accuracy while enabling fast, design-generalizable predictions. The work, accepted to ICRA 2026, promises to slash the time needed for modeling in surgical robotics and industrial automation, where quick adaptation to new robot designs is critical for control, planning, and optimization.