| Egocentric action anticipation aims to predict future actions from first-person video before they occur. However, models trained on a specific dataset often suffer significant performance degradation when deployed in visually different environments due to domain shift. In this work, we introduce EgoShift, a framework that combines unsupervised domain adaptation with reinforcement learning from human feedback (RLHF) to improve the robustness of egocentric action anticipation models in new environments. Starting from a source-domain model trained on EPIC-KITCHENS, EgoShift first adapts feature representations using unlabeled target-domain observations and then refines predictions through sparse human feedback using a PPO-based optimization strategy. To stabilize learning during continuous adaptation, the framework incorporates a replay-based mechanism that revisits previously observed samples and prevents catastrophic forgetting. We evaluate our approach on the Assembly101 dataset under cross-domain anticipation settings. Experimental results show that the proposed RLHF pipeline improves anticipation performance compared to the warm-up initialization and approaches the performance of a supervised baseline while requiring substantially less labeled data. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.