{"id":"1f414a26-7eb4-43ba-bb30-1af7dbd79e05","arxiv_id":"2509.09594","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A controller trained purely on object-relative costmaps beats image-goal navigation on alt-goal, shortcut, and reverse-route tasks, with nearly height-invariant performance and qualitative sim-to-real transfer.","lead":"ObjectReact trains a robot controller to steer using a 'WayObject Costmap', a per-object blend of segmentation masks and each object's planned distance to the goal, instead of the usual pair of current and goal images. The result is a navigation policy that keeps working when the camera height changes, when routes are reversed or shortened, and when a simulation-trained policy is deployed on a real quadruped.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sim experiments use a ground-truth submap oracle for localization; the real-world active localization is untested quantitatively, so the reported SPL gains may not survive without it.","rationale":"The reader's weakest assumption correctly points to the GT submap index as the single most consequential gap. The central claim is not merely that a costmap-conditioned controller can beat GNM when the correct map context is handed to it; the abstract claims a deployable object-relative navigation stack and real-world transfer. The sim results cannot be separated from the oracle because the oracle determines which map objects are candidates for matching and hence every path length in the costmap. The paper's own real-world active localization is a different mechanism and is only shown in videos. I agree with the reader's conditional verdict: the method is coherent, the ablations (Table 3) show genuine benefits of learned control over RoboHop/PixNav under GT perception, and the internal comparisons are not circular, but the headline numbers need a non-oracle localization test and ideally error bars/filtering transparency before the central claim can be accepted as stated.","tokens_in":19815,"tokens_out":8570,"duration_ms":106035,"concrete_test":"Re-run the Tables 1-2 evaluation in Habitat with the oracle submap removed and replaced by the paper's own active localization from §A.4: select the submap center as the map image with the most segment matches against the current query image plus its 8-frame history, then match, plan, and build costmaps exactly as in the real-world protocol. Compare ObjectReact and GNM SPL/SSPL per task and per height. If ObjectReact's margins persist, the oracle concern is resolved; if they shrink or reverse, the central quantitative claim is tied to the oracle.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Every quantitative result in Tables 1 and 2 is produced under the Section 4 protocol: \"the agent is given access to ground truth localization in the form of map image index which is closest to the agent's current 2D position.\" Concretely, §A.4 builds the WayObject Costmap by matching current FastSAM segments only against the submap centered at that oracle index (radius 16, subsampling 2). This removes the global visual localization problem entirely. The costmap is the controller's only input; if a deployable system localizes to the wrong submap, the matched objects' path lengths come from the wrong map region and the learned controller is fed corrupted costs. The real-world demos replace the oracle with active selection (submap with most matched segments over current frame + 8-frame history), which is a qualitatively harder inference problem and is never evaluated quantitatively. The paper's headline advantage on Alt Goal (21.74 vs 2.17 SPL) and the height-invariance claim in Table 2 therefore do not establish that the advantage survives without the oracle. Manual filtering of 16/36 Alt-Goal episodes (§B.2) further shrinks the sample behind the strongest numbers.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ObjectReact, a visual navigation controller that is conditioned on a 'WayObject Costmap' rather than on a pair of current and subgoal images. The map is a relative 3D scene graph built from object segments, monocular depth, and pairwise feature matching; during execution, matched objects receive global path lengths to the goal, and these are encoded as a multi-channel costmap that serves as the sole controller input. The controller is trained in HM3D and compared against an image-relative GNM baseline on four tasks: Imitate, Alt Goal, Shortcut, and Reverse. The reported results show large SPL improvements for the object-relative controller on the non-imitation tasks and much smaller sensitivity to a change in camera height between mapping and execution. Real-world deployment on a quadruped is presented qualitatively. The paper's central claim is that object-relative control decouples control prediction from image matching and prior-trajectory imitation, enabling new routes and cross-embodiment transfer.","tokens_in":20118,"tokens_out":6262,"duration_ms":81141,"significance":"If the reported results hold up, the paper makes a useful conceptual and empirical contribution: conditioning a local controller on object-level path-length costmaps is an interesting alternative to image-pair conditioning, and the height-invariance result in Table 2 is a compelling illustration of the potential benefit of object-level representations. The tasks themselves (Alt Goal, Shortcut, Reverse) are also a valuable addition to the topological navigation evaluation toolbox. The availability of code via the project page is a positive reproducibility signal. However, the strength of the evidence is currently limited by the use of a ground-truth localization oracle in all simulator experiments, by the large number of manually excluded episodes, and by the absence of uncertainty quantification. These issues directly bear on the paper's claim to demonstrate a visual navigation paradigm rather than a control policy evaluated under idealized localization.","major_comments":[{"comment":"All simulator experiments use a ground-truth localization oracle: 'the agent is given access to ground truth localization in the form of map image index which is closest to the agent's current 2D position.' The WayObject Costmap is then built by matching the current FastSAM segments only against the submap centered at that oracle index. This means Tables 1, 2, and 3 evaluate the controller conditional on perfect map context, not the full visual navigation pipeline. The real-world active localization described in Section A.4 is only demonstrated qualitatively in videos. Since the abstract and introduction claim a new 'visual navigation' paradigm and say the control prediction problem is 'decoupled from solving the image matching problem,' this protocol gap is load-bearing. Please add simulator experiments that use the active localization procedure or a controlled localization-noise model,","section":"Section 4 and Section A.4"},{"comment":"The evaluation excludes a large fraction of episodes based on manual inspection: 6/36 Imitate, 16/36 Alt-Goal, 13/36 Shortcut, and 9/36 Reverse episodes are removed. The reported Alt-Goal SPL of 21.74 vs. 2.17 and Shortcut SPL of 23.08 vs. 7.69 are therefore computed on 20 and 23 episodes, respectively. Manual post-hoc exclusion without pre-specified criteria, and without a sensitivity analysis, can bias the comparison. Please report results on all 36 episodes (for example, treating invalid episodes as failures for all methods) and provide per-episode results. If some episodes are truly invalid, the exclusion criteria should be applied identically to every method and justified in advance.","section":"Section B.2 and Tables 1-3"},{"comment":"No error bars, confidence intervals, or significance tests are reported. With 20-30 episodes per task and 72 runs total for the height experiment, the large point differences could be driven by a small number of episodes. Please report per-episode distributions, standard errors, and paired tests (e.g., Wilcoxon signed-rank) where appropriate. In addition, Table 3 is generated using ground-truth object instances from the simulator, while Table 1 uses inferred segmentation and matching; the protocol shift should be clearly flagged whenever Table 3 results are cited.","section":"Tables 1-3"},{"comment":"The GNM baseline's subgoal-image selection procedure is not specified in the evaluation. Section 5.1 attributes GNM's poor performance on Alt Goal, Shortcut, and Reverse to 'temporal distance prediction' for subgoal selection, but it is unclear whether the GNM system used its learned distance head or the ground-truth localization oracle to choose the subgoal image. This distinction matters for interpretation: an oracle-selected subgoal would give GNM an advantage on subgoal selection, while a learned distance head could introduce a separate failure source. Please specify exactly how the GNM subgoal image was selected in Tables 1-2, and ideally ablate GNM with oracle-selected subgoals versus learned-distance subgoals.","section":"Section 5.1 and Section 4"}],"minor_comments":[{"comment":"Equation (2) and the surrounding text contain a typo: 'such that l∈[1,L] Z.l=L' is garbled; presumably 'l=L represents the shortest path length per image and l=0 represents outliers.' Please clarify the encoding of the cost values.","section":"Section A.1"},{"comment":"The costmap resolution is given as W=85, H=64 in Section A.2 but as 85×65 in Section A.6. Please correct the inconsistency.","section":"Section A.2 and Section A.6"},{"comment":"Reference [61] spells the author as 'Dijksta'; it should be Dijkstra.","section":"Reference [61]"},{"comment":"There are unmatched quotation marks in '“floor” or “‘ceiling”' in the text discussing CLIP-based removal of floor/ceiling nodes.","section":"Appendix D.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is interesting and the large SPL gaps are initially compelling, but the current evaluation protocol gives the agent oracle localization in every simulator experiment and a large number of episodes are excluded after manual inspection. These issues are fixable within the manuscript's scope by adding a localization-sensitive experiment and a full-episode sensitivity analysis, or by repositioning the claims as control-policy evaluation under an oracle. The overlap with the authors' own RoboHop/TANGO line is appropriate—those are used as baselines or components—but the GNM comparison protocol needs to be spelled out. I would encourage the editor to request a revision rather than reject, because the conceptual direction and the task suite are valuable even if the current quantitative claims are stronger than the protocol supports."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi,\n\nThe short version: ObjectReact is a solid systems paper with a genuinely new conditioning representation, but the headline SPL gains all come with a ground-truth localization oracle in simulation, so treat them as controller-only results, not full-system numbers.\n\nWhat's new and good: The WayObject Costmap—per-object planned path lengths encoded into segmentation masks—is a simple, sensible idea. It removes the need for a subgoal image and lets a learned controller act without RGB. The relative 3D scene graph with Euclidean intra-image distances improves on RoboHop's 2D Delaunay triangulation, and the ablation in Table 3 backs that up. Conditioning on the costmap alone beating costmap-plus-RGB is a useful result. Using GNM as the backbone keeps the comparison clean. The appendix is transparent about perception failures, dynamic objects, and deployment gaps. No circularity: the costmap is a conditioning input, not a fitted target.\n\nWhere I'd be careful: Section 4 states every simulator run gives the agent ground-truth localization (the map image index closest to its true 2D position). The controller always starts from the correct submap. In a real system, localizing to the wrong submap corrupts the costmap, and the paper gives no quantitative evidence the advantage survives localization errors. The real-world demos use an active localizer heuristic, but they're qualitative. The manual exclusion of 16/36 Alt-Goal and 13/36 Shortcut episodes shrinks the sample behind the strongest claims. Lack of error bars is minor.\n\nThis paper is for anyone working on topological navigation, object-centric maps, or learned subgoal controllers. It is worth reading now and deserves a serious referee—I'd send it for review, while asking for results with the active localizer and a sensitivity analysis to localization error.","headline":"ObjectReact's WayObject Costmap is a real contribution, but the headline SPL gains all come with a ground-truth localization oracle, so they are controller-only numbers.","tokens_in":20617,"tokens_out":3651,"would_cite":true,"duration_ms":40249,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Object-level costmaps, not image pairs, let a robot navigate unseen routes and change heights without retraining.","keywords":["object-relative control","visual topological navigation","3D scene graph","WayObject Costmap","cross-embodiment generalization","monocular depth","segmentation-based mapping","navigation policy"],"falsifier":"In the same four simulator tasks, replace oracle localization with matcher-based localization or add random submap offsets; if SPL on Alt Goal, Shortcut, and Reverse drops toward the image-relative baseline as offset grows, the invariance claim is localization-dependent. A real-world run without any ground-truth localization would settle it.","tokens_in":19708,"feed_emoji":"🤖","tokens_out":4497,"duration_ms":54198,"temperature":0.7,"pith_summary":"ObjectReact is a navigation controller that learns from a \"WayObject Costmap\"—a per-object map of path lengths to the goal overlaid on segmentation masks—instead of from an image pair. The paper's central claim is that this object-relative representation decouples control from image matching, so the controller can handle tasks that require spatial reasoning beyond imitating a prior trajectory: reaching previously unvisited goals, cutting shortcuts, and travelling routes in reverse. On these tasks the object-relative controller scores roughly ten to twenty percentage points higher in SPL than an image-relative baseline trained on the same data, and it is almost unaffected by changing the execution camera height from 1.3m to 0.4m (SPL drop of 3 vs 48 points). The paper also reports that a policy trained only in simulation transfers to real indoor environments, and that the remaining gap on hard tasks is mostly due to imperfect perception rather than the learned controller itself.","feed_headline":"Robots steer by objects, not images, to take new routes","feed_subtitle":"A sim-only policy fed object path costs nearly ignores camera height changes and transfers to real robots.","key_machinery":"WayObject Costmap: a multi-channel image in which each pixel of an object's segmentation mask carries a sine-cosine encoded path length from that object to the goal, with unmatched or undetected segments set to a fixed outlier cost. This representation converts a variable number of object subgoals into a fixed-size learned input that retains both spatial layout and goal-directed cost, so a controller can predict waypoints without any RGB input.","core_discovery":"The central discovery is that the control problem in topological visual navigation can be re-expressed as reacting to a dense cost image built from objects. The map is a relative 3D scene graph: nodes are segmentation masks, intra-image edges are 3D Euclidean distances estimated from monocular depth, and inter-image edges link corresponding objects across frames. During execution the agent matches visible objects to map nodes, computes Dijkstra path lengths from each matched object to the long-horizon goal, normalizes those lengths per image, and encodes them with sine-cosine positional embeddings; the resulting multi-channel WayObject Costmap is the only input to a convolutional controller","pith_inferences":["Because the WayObject Costmap is a fixed-size grid encoding of goal-directed object costs, it could likely be fed to other policy classes (discrete action, diffusion) without changing the pipeline; the paper only tests its own convolutional predictor.","The closest testable extension is to add localization noise in simulation; the paper's use of oracle localization means the robustness of the object-relative advantage to real perception remains an open question.","The costmap might also be generated from language instructions or exploration objectives, extending it beyond the paper's map-based planning source and toward a more general goal-conditioning interface."],"forward_implications":["A map captured at one camera height can be executed at another height with little loss: SPL drops 3.04 points for ObjectReact versus 48.49 for the image-relative baseline.","Controller performance no longer depends on retrieving a subgoal image, so previously unvisited goals, shortcuts, and reversed routes become navigable.","Training data requirements drop from quadratic (all embodiment pairs of current and goal images) to linear, since the controller input is embodiment-invariant.","A policy trained only in simulation can generalize to real indoor environments, as demonstrated on a quadruped robot, including around obstacles that appeared after mapping."],"fun_headline_variants":["Object-relative control lets robots navigate new routes","Robots avoid image matching by steering via object costmaps","Sim-only object policy transfers to real robots for navigation","Reverse navigation works with object-relative control","Objects, not pixels: new visual navigation paradigm"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Every simulator run hands the agent the correct map image index as localization, so the costmap is always built from the right map context; if localization is noisy, the object-relative advantage could shrink or vanish.","fun_headline_variants_meta":{"raw":{"variants":["Object-relative control lets robots navigate new routes","Robots avoid image matching by steering via object costmaps","Sim-only object policy transfers to real robots for navigation","Reverse navigation works with object-relative control","Objects, not pixels: new visual navigation paradigm"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00018,"raw_usage":{"total_tokens":1177,"prompt_tokens":818,"completion_tokens":359,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":286}},"tokens_in":562,"tokens_out":359,"duration_ms":4667,"temperature":1.0,"reasoning_tokens":286,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T18:48:49.170410+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the same four simulator tasks, replace oracle localization with matcher-based localization or add random submap offsets; if SPL on Alt Goal, Shortcut, and Reverse drops toward the image-relative baseline as offset grows, the invariance claim is localization-dependent. A real-world run without any ground-truth localization would settle it.","supporting_citations":[],"review_version":1}