Data-Centric Recursive Self-Improvement for Navigation via Policy-Synchronized Language-Action Dual Bootstrapping
RSI-Nav updates its own generated navigation data shard by shard, so supervision does not fall behind the policy being trained.
The paper proposes one shared model for both following navigation instructions and generating instructions from trajectories. After supervised fine-tuning, it labels small batches of unlabeled trajectories, tests whether the resulting instructions lead to efficient goal completion, keeps the qualified samples, and immediately trains on them before moving on. The authors add spatial-understanding data to support navigation reasoning. They report 68.6% success rate on R2R-CE and 72.4% on RxR-CE, plus stronger instruction-generation utility after recursive evolution. HF Daily Papers' note
The paper proposes one shared model for both following navigation instructions and generating instructions from trajectories. After supervised fine-tuning, it labels small batches of unlabeled trajectories, tests whether the resulting instructions lead to efficient goal completion, keeps the qualified samples, and immediately trains on them before moving on. The authors add spatial-understanding data to support navigation reasoning. They report 68.6% success rate on R2R-CE and 72.4% on RxR-CE, plus stronger instruction-generation utility after recursive evolution. HF Daily Papers' note
score 4