Failure Modes of Zero-Shot Machine Unlearning in Reinforcement Learning and Robotics

Machine unlearning, the targeted removal of data influence from trained models, is becoming increasingly important as intelligent robotic systems operate in dynamic environments and must comply with evolving privacy regulations, such as the General Data Protection Regulation (GDPR) and its "right to be forgotten" provision [4].Zero-shot unlearning has emerged as a promising approach, enabling robots and reinforcement learning (RL) agents to forget specific data or classes without retaining the original training data, which is essential for flexible deployment and adaptation.In this work, we highlight a key limitation of current zero-shot unlearning frameworks.We find that the accuracy on the forget set, after initially dropping, can unexpectedly recover with continued training, indicating that forgotten information may gradually resurface.This insight highlights the need to advance robust unlearning techniques in both image processing and robotics, particularly methods that can securely and efficiently unlearn not only data but also environment dynamics.Addressing these challenges will be crucial for building robotic systems that can safely and reliably adapt to new environments and comply with emerging regulatory requirements such as the AI Act [5] in the future.