Precise terminal positioning and heading alignment remain hard for learning-based mobile robot control. While vision-based methods with pretrained encoders generalize to new scenes, general-purpose visual representations often miss the fine-grained spatial differences between current and goal views needed for accurate control. The authors of GR-VIP argue that this gap limits real-world demonstrations that combine millimeter-scale accuracy with zero-shot transfer.

GR-VIP is a goal-conditioned visuomotor policy built on a visual encoder pretrained for relative camera pose estimation. Instead of using the final relative-pose estimate, the policy conditions action prediction on the dense relational features produced before pose regression. This preserves cross-view spatial information that helps the policy learn corrective actions from demonstrations.

The method was evaluated on a real-world docking task. It achieved high docking success rates and millimeter-scale terminal accuracy in both the training environment and unseen environments, with no additional training for transfer. Supplementary experiments also showed the policy could align with non-docking targets and complete docking under a tested low-light condition.

The paper is a single source and has been submitted to IEEE Robotics and Automation Letters; it has not yet been peer-reviewed. The reported results are promising but independent replication and further testing would be needed to confirm the method's general applicability beyond the specific docking scenarios described.