OpenGVLab releases RSI-Nav, an open-source recursive self-improvement framework for vision-language navigation
OpenGVLab released RSI-Nav, an open-source recursive self-improvement framework for continuous vision-language navigation built on Qwen3.5-9B. The repository provides training, trajectory conversion, self-improvement, and R2R-CE evaluation code. Model weights and datasets are prepared separately.
China context
- Original name
- 上海人工智能实验室
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders can use the RSI-Nav codebase to implement recursive self-improvement for navigation models, but must wait for model weights and datasets to be released separately.
- For investors
- The release signals OpenGVLab's continued investment in embodied AI research, but the lack of weights and datasets limits immediate commercial applicability.
RSI-Nav is a data-centric recursive self-improvement framework for continuous vision-language navigation, built on Qwen3.5-9B. It improves instruction following and generation through iterative instruction generation, navigation verification, and joint training with spatial-data replay. The repository includes code for DE-to-CE conversion, supervised fine-tuning, recursive self-improvement, and R2R-CE evaluation, supporting multi-node, multi-GPU execution. The reference environment specifies Python 3.11.12, PyTorch 2.7.0+cu128, Transformers 5.2.0, Habitat-Sim 0.3.3, and FlashAttention 2.8.3.
RSI-Nav employs policy-synchronized language-action dual bootstrapping, combining instruction generation, navigation verification, and joint training with spatial-data replay. It is built on Qwen3.5-9B and integrates with Habitat-Sim for embodied navigation. The framework supports multi-node, multi-GPU training and evaluation on R2R-CE.
Developers outside China gain access to a complete open-source codebase for recursive self-improvement in vision-language navigation, reducing the cost of replicating state-of-the-art embodied AI research. Competitors can now benchmark their own navigation models against RSI-Nav's R2R-CE evaluation setup.
RSI-Nav provides a foundation for building navigation systems that improve through self-generated data, potentially lowering data annotation costs for robotics and embodied AI applications.
The release of model weights and datasets is pending; their availability will determine whether RSI-Nav can be fully reproduced. The next verifiable signal is the publication of the associated paper or the release of pretrained checkpoints.