- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This work introduces World SLAM Model, an approach that unifies SLAM and navigation, by bridging world modeling of future states with traditional SLAM methods. In a nutshell, the approach uses video generation to predict (dream) future frames towards a goal and then SLAM processes these frames to produce future poses and geometry (essentially working as an inverse dynamic model). These poses are also used for RL-based post training. The core of the approach uses a Mixture of Transformers to jointly train future prediction (generation expert) and geometric prediction (SLAM expert). The approach is evaluated on a navigation benchmark and a SLAM benchmark.
Critical strengths
- Connecting world models and SLAM systems is a reasonable and seemingly technically sound approach that could be expected to improve navigation performance.
- Leveraging the reconstructed geometry for RL self-training seems a clear advantage of the proposed framework.
- The approach is shown to outperform baselines in experiments, and the ablations support the design choices.
Critical weaknesses
- Using the term “SLAM” in this context seems misleading since the approach does not retain a persistent memory and does not perform loop closure. Technically, this provides a long context, but it cannot maintain the full history of information throughout execution. What happens if we gather more than 128 latent frames? How does performance change as the scene gets larger? The experiments seem to focus on relatively short trajectories (Figure 4).
- The SLAM evaluation is not comprehensive enough. The authors should examine other SLAM methods, such as VGGT-Omega, ORB-SLAM, and DROID-SLAM, and evaluate on datasets such as TUM, Replica, 7-Scenes, KITTI, or similar if they are claiming a contribution to SLAM.
- The paper does not fully justify why "joint modeling" or "joint training" is necessary or why it should be considered a contribution. I suggest comparing against a model variant where the generation expert and SLAM expert are not jointly trained but are connected in a modular manner. For example, you could train the video generation model to condition on pose and depth, and use a state-of-the-art dense 3D prediction model, such as VGGT to estimate pose and depth. If your jointly trained model performs better than a separately trained but serially connected model, then the ablation would better justify the proposed design.
- Limited navigation benchmarks. InternNav includes two datasets (https://github.com/InternRobotics/InternNav). The paper reports performance on InternScene, but where are the results for the ClutteredEnv dataset? Since both datasets are under the same codebase, I assume it should be easy to evaluate on the latter. For more comprehensive experiments, the authors should also consider other simulation environments, such as Habitat and Isaac Sim. Only 40 scenes from InternScene do not provide sufficient evidence of the performance of the proposed approach.
- Navigation baselines. The paper presents the use of a world model as a core module for planning as one of its contributions. However, the chosen baselines are all planners without a world modeling component. This raises concerns about whether the proposed method is actually at the state-of-the-art in world-model-based navigation. The authors should consider comparing against Navigation World Models [1]. To make a convincing argument that "joint modeling" contributes to performance, the authors should consider comparisons against modular navigation methods with a world modeling or future imagination component, such as STRIVE [2], SGImagineNav [3], and ImagineNav [4]. In addition, the future prediction quality is not evaluated. How does the quality of predicted frames affect pose and depth estimation? How does it affect navigation performance?
- No video attachment showing a full navigation episode. The paper only shows the final results and snapshots of navigation. However, it is difficult to be convinced that the method works smoothly without a full video showing a successful navigation episode. In addition, the dreamed visuals seem to be very short-horizon. With such local predictions, how can the system avoid following suboptimal trajectories and getting stuck in places such as tight corners?
[1] Bar, Amir, et al. "Navigation World Models." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025 [2] Zhu, Hao-Kun et al. “STRIVE: Structured Representation Integrating VLM Reasoning for Efficient Object Navigation.” ArXiv abs/2505.06729 (2025) [3] Y. Hu et al., "Imaginative World Modeling With Scene Graphs via VLM for Object Navigation," in IEEE Robotics and Automation Letters, vol. 11, no. 11, pp. 12855-12862, Nov. 2026, doi: 10.1109/LRA.2026.3732899. [4] Zhao, Xinxin, et al. "ImagineNav: Prompting Vision-Language Models as Embodied Navigator through Scene Imagination." International Conference on Learning Representations, 2025
Questions for authors
- The paper does not fully justify why "joint modeling" or "joint training" is necessary or why it should be considered a contribution. How does the proposed approach compare against a model variant where the generation expert and SLAM expert are not jointly trained but are connected in a modular manner?
- What are the runtime and memory requirements of the model? Although the paper mentions that inference runs on an RTX 4090 GPU, providing runtime measurements would offer useful insights into practical deployment.