GAP: Geometric Anchor Pre-training Boosts Robot Learning from Few Demonstrations

A new approach called Geometric Anchor Pre-training (GAP) has been introduced by researchers to enhance data efficiency in robotic manipulation through visuomotor policy learning. GAP tackles the issue of limited expert demonstrations by regularizing a spatial pooling module that adjusts pre-trained Vision Foundation Models (VFMs) for control tasks. In the absence of GAP, this adapter risks overfitting to irrelevant shortcuts and may lose its geometric foundation when fine-tuned with minimal samples. Prior to downstream imitation learning, GAP pre-trains the pooling layer using a lightweight simulated proxy task, thereby incorporating robustness-oriented inductive biases. This straightforward, action-free method acts as a preparatory phase. The research is available on arXiv with ID 2605.15836.