NeurIPS paper learns compositional motor options with adapter banks, cutting generalization error up to 10x

sreejan_kumar · x · 2026-09-25

A paper by Sreejan Kumar, Marcelo Mattar and Lea Duncker accepted to NeurIPS 2026 translates the neuroscience idea that motor primitives are low-rank perturbations of a shared recurrent network into an end-to-end architecture: a shared recurrent core modulated by a bank of residual adapters, each selected via a discrete latent code.

The author calls it a precursor to what he hopes will be one of the biggest papers of his career.

Original post →

More from Embodied

Embodied channel →