Soft robots just got a much faster brain for simulation and control.
Researchers released SoRoMoX, a Python framework built on JAX that models soft robots using Cosserat-rod theory, the standard math for simulating flexible, continuum-like limbs. Unlike older tools, it runs on GPUs and is fully differentiable, meaning every force, torque, and parameter in a simulation can feed directly into an optimization or learning loop. On a single CPU, SoRoMoX's rollouts run up to 27 times faster than SoRoSim, an established comparison framework; parallelized on a GPU, that speedup climbs to as much as 679.7 times for its general variable-strain model. The same differentiability also improved related tasks: system identification saw 66% lower marker error, computed-torque tracking cut error by roughly 500 times versus basic PD control, and safety-constrained control held contact forces under a 5 N limit instead of 33.5 N without that constraint.
Soft robotics has had workable models for years - the real bottleneck was turning those models into something fast enough to actually train on. SoRoMoX's biggest practical win may be reinforcement learning (RL): training a control policy ran up to 7 times faster than a CPU-based baseline called PyElastica, thanks to massively parallel GPU rollouts. That is the difference between a multi-day training run and an afternoon one.
Rigid robots got GPU-accelerated, differentiable physics years ago; soft robots, squishier and harder to model, are only now catching up.