A new algorithm claims something rare in evolutionary robotics: a guarantee that it found the best design, not just a good one.
The claim comes from a paper posted to arXiv on September 25, 2026 (arXiv:2609.29491, https://arxiv.org/abs/2609.29491). It introduces MISCO, a framework for designing voxel-based soft robots - modular, cube-built creatures that move by deforming rather than using rigid joints. MISCO pairs an estimation-of-distribution algorithm with a customized variational autoencoder that tracks voxel position and lets neighboring voxels signal each other, then uses that combination to search the design space more efficiently. The paper includes a mathematical proof that the method converges to the globally optimal design, plus simulated tests across multiple tasks.
That matters because soft robot design has largely been a brute-force problem. The space of possible voxel arrangements is enormous, and testing each one - even in simulation - is slow, so most evolutionary search methods settle for local optima and call it a day. A framework with a convergence proof, even a theoretical one, gives engineers a way to know whether more search time is worth spending or whether they have already hit the ceiling.
Worth noting: the guarantees and results here are asymptotic and simulated. No physical soft robot has been built from a MISCO design yet, and simulation-to-reality gaps have derailed plenty of promising robotics papers before this one.