We propose View Invariant Learning (VIL), a strategy that adapts existing navigation policies to varied viewpoints without retraining from scratch. VIL contains two components: a contrastive learning objective and a teacher-student framework for waypoint prediction.
The contrastive component encourages sparse and view-invariant representations by aligning observations of the same scene captured from different viewpoints while separating unrelated observations. Features for contrastive learning are produced by a projection head and are shared with the navigation policy.
For waypoint prediction, a frozen teacher model, initialized from a pretrained navigation policy, processes observations from a standard viewpoint. The student model shares the same architecture but trains only a lightweight adapter inserted into the waypoint predictor while keeping the remaining parameters frozen. The student receives varied-viewpoint observations and learns to match the teacher outputs through a distillation loss.
Both components are optimized jointly in an end-to-end manner, enabling efficient adaptation to viewpoint variation while preserving the original navigation capability.
