Beyond Mimicry: Toward Lifelong Adaptability in Imitation Learning
Nathan Gavenski, Felipe Meneguzzi, Odinaldo Rodrigues
TL;DR
Imitation learning currently prioritizes replay fidelity over adaptability, causing brittle performance when contexts or goals shift. The authors propose Compositional Repertoire Learning (CRL), which aims to extract reusable primitives and compositional rules from demonstrations and recombine them in novel contexts without retraining. They formalize Goal-conditioned Contextual MDPs (GCMDPs), define compositional generalisation dimensions (systematicity, productivity, substitutivity), and introduce evaluation frameworks based on controllable contexts and deterministic transitions, including a generalisation boundary concept driven by a maximum contextual distance $δ_C$. The work argues for a shift toward lifelong adaptability, discusses benchmarks and future research directions, and envisions integrating IL with planning and foundation models to enable open-ended, robust, and explainable autonomous agents.
Abstract
Imitation learning stands at a crossroads: despite decades of progress, current imitation learning agents remain sophisticated memorisation machines, excelling at replay but failing when contexts shift or goals evolve. This paper argues that this failure is not technical but foundational: imitation learning has been optimised for the wrong objective. We propose a research agenda that redefines success from perfect replay to compositional adaptability. Such adaptability hinges on learning behavioural primitives once and recombining them through novel contexts without retraining. We establish metrics for compositional generalisation, propose hybrid architectures, and outline interdisciplinary research directions drawing on cognitive science and cultural evolution. Agents that embed adaptability at the core of imitation learning thus have an essential capability for operating in an open-ended world.