Paper review · October 15, 2026
RSI-Nav: Recursive Self-Improvement for Navigation via Policy-Synchronized Language-Action Mutual Bootstrapping
Data-Centric Recursive Self-Improvement for Navigation via Policy-Synchronized Language-Action Dual Bootstrapping
A single model generates, executes and verifies route instructions itself and immediately updates itself after each small data chunk (shard). This recursive self-improvement raised the VLN-CE success rate.
Problem
Vision-language navigation (VLN) needs a lot of instruction data, but human annotation is expensive. Existing synthesis methods keep the instruction generation model and the navigation model separate or use task-specific architectures. Moreover, they update the policy only after annotating the whole corpus with a fixed model, so the generated supervision signal lags behind the evolving policy.
Method
A single Qwen3.5-9B handles instruction following (predicting actions in units of 4), instruction generation (taking 32 frames and an action sequence as input) and spatial understanding. It first gains basic abilities through SFT. Then unannotated trajectory data is split into small shards; for each shard, the current model generates instructions and executes them itself from the starting point. Only pairs that both reach within 3 m of the goal and meet an SPL threshold are kept. The kept pairs are matched with the original trajectories and used as training data in both directions, instruction following and generation, mixed again with spatial understanding data (replay), and the model is updated immediately before moving to the next shard.
Results
- Success rate (SR) of 68.6% on R2R-CE Val-Unseen. Higher than the waypoint-based models in the same table, GridMM (49.0%) and VLN↻BERT (44.0%)
- Success rate of 72.4% on RxR-CE Val-Unseen. Higher than VLN↻BERT (27.0%) in the same table
- The recursive improvement process produced about 2.5 million execution-verified instruction-following samples and about 1.1 million instruction generation samples
- They report small improvements over the initial SFT model on 4 spatial understanding benchmarks (VSI-Bench, ScanQA, SQA3D, MindCube)
Limitations
The source trajectories are fixed offline data, so this is not strictly on-policy training (the authors call it 'policy-synchronized'). Because the original trajectories, not the executed ones, serve as the supervision signal, the improvement targets mainly language supervision. Evaluation is limited to Habitat-based simulation environments. The provided text was cut off, so detailed comparisons with the latest baselines and ablation numbers could not be checked.
Why it matters
It shows a loop in which the model verifies self-generated data through execution and feeds it straight into training. It is a useful reference for data-centric self-improvement in embodied AI, where human annotation is expensive.
Key terms
- VLN-CE
- Vision-language navigation in continuous environments. The agent follows natural-language instructions and moves to the goal using low-level actions such as moving forward and turning.
- SPL(Success weighted by Path Length)
- A navigation efficiency metric that reflects both whether the goal was reached and how close the path is to the shortest path.
- Recursive self-improvement (RSI)
- An iterative process in which a model creates and filters the training data for its next version, and the improved model in turn creates better data.
Original
Authors: Songze Li, Haoming Gu, Ziyang Gong, Zhitong Wang, Gengze Zhou, Zun Wang, Yuhan Zhu, Tianxiang Jiang, et al. · Posted 2026-10-11 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.