Robot policies usually map observations to actions with multilayer perceptrons (MLPs). But robot dynamics and control objectives often involve multiplicative interactions among observable physical variables. The arXiv abstract notes that MLPs can approximate these interactions, though the paper argues for a more structured approach.

The paper introduces PRISM, which stands for Polynomial Representations for Interaction-Structured Motor Control. Instead of relying solely on generic MLPs, PRISM uses polynomial representations designed to explicitly model the multiplicative structure inherent in robot dynamics and control.

The abstract is brief and does not include experimental results or comparisons. It is available on arXiv under identifier 2607.23473v2, and the full paper may provide further details on the proposed method and its performance.