World-Task Factorization for Robot Learning

Abstract

A framework that factors robot policies into world and task components, motivated by a Bayesian model-evidence argument. The world factor reuses AICON, a compositional, differentiable graph of recursive estimators that propagates cost gradients to actuators, and a compact learned policy modulates the resulting gradient paths to resolve task trade-offs. Across heterogeneous multi-robot search, bimanual handover, and pressure-plate tasks, the approach outperforms end-to-end learned baselines, generalizes zero-shot to out-of-distribution configurations, and transfers to real robots without retraining.

Paper