Contextual Forgetting: Mitigating Knowledge Obsolescence for Safe Lifelong Robot Learning

ICLR 2026 Conference Submission17002 Authors

Keywords: Lifelong Robot Learning, Robot Safety, Knowledge Obsolescence

Abstract: Lifelong Robot Learning, in its pursuit of general intelligence, confronts a critical yet overlooked challenge: endogenous safety risks arising from "knowledge obsolescence." When a once-optimal policy becomes detrimental after an environmental shift, the conventional Continual Learning (CL) paradigm, which focuses on "remembering," lacks an active "forgetting" mechanism, posing significant risks in the physical world. To address this, we introduce "Contextual Forgetting," a novel mechanism, and design a Knowledge Validity Module (KVM). The core of KVM is a principled risk assessment framework based on an Energy-Based Model (EBM), enabling it to actively identify and mitigate hazardous interactions caused by knowledge inapplicability. We validate the efficacy of this framework by deeply integrating it with CODA-Prompt, an advanced CL algorithm. Experiments demonstrate that KVM significantly reduces catastrophic failures caused by knowledge obsolescence without sacrificing learning efficiency, providing a rigorous solution for building safer and more reliable lifelong learning robotic systems.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 17002