DeepSoMo estimates how soft robots deform from their own internal sensing

Researchers led by Hanwen Zhao have presented DeepSoMo, a learning framework for estimating how deformable bodies in soft robotics move from proprioceptive sensing, in a paper posted on arXiv.

Internal sensors give only indirect clues about overall motion, contact and shape, so single readings are not enough, the authors write. They argue that long estimates need latent states that stay predictive of future observations, and DeepSoMo enforces this with a self-supervised latent prediction constraint.

Tested in a standard deformable-dynamics setting with high-fidelity finite-element simulation as ground truth, it estimated trajectories, orientation, shape and contact more accurately than recurrent baselines with matching architectures, the authors report.