SoRoMoX brings 234x faster GPU-accelerated soft robot simulation

SoRoMoX (Soft Robot Models in JAX) is a new open-source framework from researchers at MIT, TU Delft, and other institutions that finally brings soft-robot simulation into the modern machine-learning era. Unlike existing reduced-order models based on Cosserat-rod theory—which are accurate but slow and non-differentiable—SoRoMoX is JIT-compilable and runs entirely on GPUs. It implements articulated, Piecewise Constant Strain, and Variable Strain models through a unified interface, with inertia matrices, Jacobians, and derivatives all end-to-end differentiable with respect to states, inputs, and parameters.

Performance is dramatic: sequential CPU rollouts are up to 18.1x faster than state-of-the-art alternatives, while GPU-parallel rollouts boost throughput by up to 234.6x. That speed unlock enables previously impractical workflows. Static-equilibrium system identification achieves 66% lower marker error, residual-force learning adds another 64% improvement, and computed-torque control cuts tracking error by roughly 500x compared to model-free PD. Safety-constrained control using high-order control barrier functions keeps peak contact force within a strict 5 N limit (vs. 33.5 N without), and reinforcement-learning policy training runs up to 7x faster than a CPU PyElastica baseline.

The team positions SoRoMoX as the missing bridge between soft-robot theory and the control workflows that already power rigid robotics—think model-predictive control, reinforcement learning, and online parameter adaptation. Because the entire simulation stack is differentiable and GPU-native, researchers can train policies in minutes rather than hours, optimize controller gains automatically with up to 62% lower loss, and identify plant models from real sensor data far more reliably.

SoRoMoX is built on JAX, making it immediately compatible with existing deep-learning ecosystems and easy to deploy across research labs and industry R&D teams. The code is available on arXiv (2608.06650), and the team expects registration as a DOI soon. For anyone working on soft grippers, surgical robots, or bio-inspired actuators, this framework converts simulation from a bottleneck into an accelerator.