# Navigation robots learn to think before they move—could occlusion finally be solved? Zero-shot navigation: Schrödinger's Navigator Full article: https://haiguangboy.com/posts/schrodinger-navigator ## Core method # Navigation robots learn to think before they move—could occlusion finally be solved? ## Key results ## One-sentence take Zero-shot navigation: Schrödinger's Navigator ## What problem does the paper address Where is zero-shot object navigation (ZSON) stuck? Not semantic understanding—pure semantic baselines like InstructNav already hit 22/30 in static scenes. The real bottleneck is "myopia": how to locate targets in occluded areas, how to handle moving obstacles appearing, the current visual input offers no answers. ✅ Solution: make navigation robots "think before they move" ↳ ① Tri-trajectory sampling: generate three candidate trajectories per step (obstacle-avoiding/exploration/goal-directed), covering different decision spaces ↳ ② FlashWorld imagining the future: call a 3DGS world model for each trajectory, generating geometrically consistent future scenes in seconds ↳ ③ Future-Aware Value Map decision: blend current observation value map m and imagined future value map m_FA (weight β), unifying semantic goal guidance, information gain, and safety to select the optimal trajectory Validated on a Unitree Go2 quadruped robot across three indoor scenes: 27/30 vs InstructNav 22/30, with gains driven by dynamic targets and sudden obstacles. ## Core method ## Experimental results ## What this paper is not The gains are conditional: in static, tidy environments it matches baselines, and each step still pays a fixed cost of three trajectories × three world model inferences, with no adaptive computation allocation. More practical issue: FlashWorld only says generation time is "within seconds," without specific latency numbers, and the Go2's onboard compute configuration is undisclosed. Whether an Orin NX can run at a 15fps navigation frequency is a hard question that must be tested before deployment. ## Implications for robotics / world models / VLA This paper directly clashes with WAM in the library—WAM says "future imagination at inference is useless; the key is video-coordinated objectives during training"; Schrödinger's Navigator says "in occluded/dynamic scenes, inference-time imagination is the core gain." The two papers' 3/3 strong consensus contradicts each other, suggesting it's not either side's fault but scenario-dependent: the stronger the dynamic occlusion, the more valuable inference-time imagination; in static environments, training-time representations are the main variable. Meanwhile, D4RT focuses on pixel-level reconstruction of dynamic scenes, QWEN-RobotWorld builds state-transition world models—three papers from reconstruction/prediction/planning angles point to the same direction: dynamic perception is the next real frontier. For nav-brain's Orin NX real-time challenges, refer to π0.7's RTC asynchronous inference engineering solution. ## Key pages of the paper ![paper page](https://static.haiguangboy.com/papers/schrodinger-navigator/page_01.webp) ![key figure](https://static.haiguangboy.com/papers/schrodinger-navigator/key_figure.webp) ## Related route comparison - Opposite route · [No need to imagine the future at inference—robots still hit 91.8%! Fast-WAM debunks WAM's core assumption](https://haiguangboy.com/posts/fast-wam) `fast-wam_2026_07`: training video objectives matter more than imagining the future at test time - Same route · [D4RT_Efficiently_Reconstructing_Dynamic_Scenes_One](https://haiguangboy.com/posts/d4rt_efficiently_reconstructing_dynamic_scenes_one) `d4rt_efficiently_reconstructing_dynamic_scenes_one_2026_06`: ZSON's real bottleneck is occlusion and dynamics, not pure semantic understanding—the myopia of current visual input is the core gap validates D4RT using a single feedforward interface to output point clouds/depth/point trajectories/camera parameters (static+dynamic) from video ## Boundaries ## Author's judgment (not paper content, cross-paper synthesis) This paper directly clashes with WAM in the library—WAM says "future imagination at inference is useless; the key is video-coordinated objectives during training"; Schrödinger's Navigator says "in occluded/dynamic scenes, inference-time imagination is the core gain." The two papers' 3/3 strong consensus contradicts each other, suggesting it's not either side's fault but scenario-dependent: the stronger the dynamic occlusion, the more valuable inference-time imagination; in static environments, training-time representations are the main variable. Meanwhile, D4RT focuses on pixel-level reconstruction of dynamic scenes, QWEN-RobotWorld builds state-transition world models—three papers from reconstruction/prediction/planning angles point to the same direction: dynamic perception is the next real frontier. For nav-brain's Orin NX real-time challenges, refer to π0.7's RTC asynchronous inference engineering solution.