{"id":"9f034480-64e2-4c81-995e-87c866ddee0a","arxiv_id":"2509.08699","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A navigation pipeline that bridges object-level global planning with traversability-aware local control, using only RGB images and pretrained models, improves success over prior zero-shot and learned baselines in simulation.","lead":"A robot navigation system combines a global map of objects with local obstacle-avoiding control using camera depth and floor detection, with no pre-built 3D map. It shows a modular, zero-shot approach using off-the-shelf AI models that can reach new goals indoors, relevant to home robots and embodied AI.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Monocular depth from Depth-Anything is treated as metric without calibration; the central 'metric local control' claim rests on this unverified assumption.","rationale":"The reader's weakest assumption identifies the same issue: monocular depth from Depth-Anything is used as metric depth without calibration. This is the most load-bearing concern because the paper's headline contribution is the integration of metric local trajectory planning with topological global planning. If the depth scale is arbitrary, the BEV traversability map is not metric, and the claimed advantage over RoboHop's purely topological servoing is not attributable to metric planning. The paper's own ablation partially mitigates the concern but does not isolate depth scale or calibrate it. The verdict should remain conditional: the method may work, but the central claim depends on an assumption that is neither stated nor tested. I do not see an internal inconsistency or a fraudulent result; the issue is an omitted calibration/demonstration. Therefore no change to the reader's CONDITIONAL verdict is needed.","tokens_in":14150,"tokens_out":10912,"duration_ms":123945,"concrete_test":"On the same 108 HM3D/IIN episodes, run TANGO's No-GT controller in three depth configurations with all other perception/planning unchanged: (1) Depth-Anything output used exactly as in the paper; (2) the same output rescaled by the per-frame scale that best fits the simulator's ground-truth metric depth (or, for real-world, by a median-depth calibration to known camera height); (3) the same output multiplied by fixed global scales of 0.5 and 2.0. Report easy/hard/full success rates with confidence intervals. If (1)≈(2), the metric-scale ambiguity is not load-bearing and the paper only needs a clarification; if (1) differs from (2), the central claim depends on an unstated calibration and the 'metric' local control claim is not yet demonstrated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-B states: 'we apply monocular depth estimation via Depth-Anything [14] to project the traversable segments and sub-goal points into 3D space, resulting in the final metric BEV traversability map.' Depth-Anything's released model outputs affine-invariant/relative depth, not metric depth. The paper provides no calibration, scale factor, or uncertainty model. The local controller selects the 'farthest projected point' of the sub-goal segment and runs Dijkstra on a BEV cost map built from this projection. If the per-image depth scale and shift are arbitrary, the BEV map is not metric: obstacle sizes, distances, and the selected sub-goal point are all in an unknown, possibly frame-varying coordinate system. This is load-bearing because the central claim is precisely that TANGO adds metric local control on top of topological planning. Table II's comparison of simulated depth+traversability versus Depth-Anything+FastSAM shows only a ~5% drop on 'hard' episodes, but it changes two components at once and uses a binary success metric on 3–5m trajectories; it does not establish cross-view metric consistency. If depth scale is wrong, the reported success may come from the yaw-only fallback controller or from the intrinsic robustness of the topological sub-goal signal, rather than from the claimed metric BEV planning.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents TANGO, a zero-shot RGB-only navigation pipeline that couples RoboHop's object-level topological graph planning with a local BEV traversability controller. At each step, SAM segments are assigned topological costs; CLIP filters traversable segments; Depth-Anything projects them into a BEV map; Dijkstra plans a path to the farthest point of the selected sub-goal segment; and an auto-switch falls back to RoboHop's segment servoing when no traversable region is detected. Experiments on 108 HM3D InstanceImageNav episodes across easy/hard/full trajectory lengths, with GT-Metric and GT-Topological ablations, compare TANGO against RoboHop and PixNav, plus a seen-but-unvisited goal evaluation and real-world demonstrations.","tokens_in":14504,"tokens_out":5334,"duration_ms":53766,"significance":"If the reported results are reproducible, TANGO is a useful modular alternative to end-to-end learned navigation: it is open-set, interpretable, and zero-shot, and the code release facilitates verification. The strongest evidence is the consistent improvement over RoboHop in the No-GT setting (Table I), and the seen-but-unvisited goal results (Table IV) extend teach-and-repeat beyond visited poses. However, the central 'metric local control' claim rests on an uncalibrated monocular depth source, and the performance differences are reported without uncertainty, so the current evidence is suggestive rather than conclusive.","major_comments":[{"comment":"The paper refers to 'monocular metric depth' and a 'metric BEV traversability map' (Section III-B) and the central claim in Section I is that the robot 'metrically plans its motion'. However, the cited Depth-Anything (CVPR 2024) outputs affine-invariant, not metric, depth, and no calibration, scale factor, or uncertainty model is given. Since the controller selects the farthest projected point of the sub-goal segment and runs Dijkstra on the BEV cost map, unknown per-frame scale/shift can distort distances, obstacle extents, and the selected sub-goal point. Table II confounds changes in depth and traversability and uses a binary success metric, so it does not establish metric consistency. Please either add a scale calibration (e.g., camera height against a known floor plane) or explicitly reframe the controller as using relative depth and analyze what 'metric' then means.","section":"III-B"},{"comment":"All success rates in Table I are single point estimates from 108 episodes with no confidence intervals, error bars, or significance tests. In the No-GT setting, the hard-trajectory gap between TANGO (43.14) and RoboHop (34.56) is about 8.6 percentage points; with 108 episodes, the approximate binomial standard error is around 5 points, so this difference may not be significant. The full-trajectory gap (21.57 vs 13.73) is even less certain. Please report the number of successful episodes per cell, confidence intervals, or per-episode results, and perform a statistical comparison to support the headline outperformance claim.","section":"Table I"},{"comment":"The PixNav baseline is given a non-standard protocol: it receives topologically planned sub-goals and updates them only when the model outputs 'Done' or its memory buffer is full, while TANGO and RoboHop use a continuous, fixed-camera controller. The paper acknowledges this in Section IV-B, but the abstract claims outperformance against 'existing state-of-the-art methods' without this qualification. The comparison against RoboHop is unaffected, but the PixNav result should be either made more faithful to PixNav's intended usage or explicitly framed as a protocol-specific comparison.","section":"IV-B and V-A"}],"minor_comments":[{"comment":"Replace 'monocular metric depth' with 'monocular depth' or describe the calibration procedure; the current phrasing overstates what Depth-Anything provides.","section":"Abstract and III-B"},{"comment":"The caption should state the setting (GT-metric, 250 steps) explicitly; the 'Improvement' column gives 11.64 but the table lists 62.14 and 73.78, so the meaning should be clarified.","section":"Table III"},{"comment":"The real-world tests are only mentioned as supplementary video. Add at least the robot platform, number of runs, and quantitative outcomes (success/failure and any failure modes).","section":"V-D"},{"comment":"Typos: 'Go als' in the title and 'contirbution' in the author footnote.","section":"Title page"},{"comment":"The limitations paragraph lists perception, planning, and traversability errors, but not the monocular depth scale assumption. Please add it to the limitation list, or justify why it is not a limitation.","section":"VI"}],"recommendation":"major_revision","confidential_remarks":"The paper is built closely on RoboHop, a prior publication by four of the authors; the incremental contribution is the local metric controller. The uncalibrated depth issue is the main technical risk and could be addressed with a simple floor-plane scale estimate; the lack of error bars is also fixable. I believe the paper is repairable within its scope, provided the authors either calibrate the depth or carefully re-scope the 'metric' claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: TANGO is a genuinely new combination—object-level topological planning (RoboHop's graph) feeding a traversability-aware BEV local planner built from monocular depth and CLIP-based segment filtering, with an auto-switch fallback to topological segment servoing. That integration is not in the prior literature: RoboHop does pure segment servoing, PixNav is a learned policy. The paper also releases code and demonstrates a seen-but-unvisited goals capability beyond teach-and-repeat, which is a nice step. So there is real substance here.\n\nThe soft spots are mostly about evidence and one load-bearing assumption. The biggest is the metric depth question. Section III-B says Depth-Anything projects traversable segments into 'the final metric BEV traversability map.' Depth-Anything (v1, which is what's cited) outputs affine-invariant depth; unless you calibrate scale/shift per scene or per frame, the BEV cost map is not metric in any meaningful sense. The paper gives no calibration step, no uncertainty model, no cross-view consistency check. That matters because the claim is precisely that TANGO adds metric local control. The ablation in Table II—only a 5% drop from simulator depth to Depth-Anything on hard episodes—suggests the system is robust to this, or that the fallback controller is masking failures, or that the topological sub-goal signal is doing more of the work than the metric planner. Any of those weakens the title claim. The stress-test note is on target.\n\nOther issues, in order of size: all success rates are single point estimates with no error bars, on 108 episodes; the PixNav baseline is given a non-standard protocol (intermediate goals injected, different action space), so the 'state-of-the-art' framing in the abstract exceeds what a two-baseline comparison supports; Table III changes the step budget and setting, so the auto-switch gain is not directly comparable; real-world tests are described but not quantified. None of these is fatal, but together they mean the empirical case is conditional, not definitive.\n\nThe citation pattern is acceptable: RoboHop is self-cited but it's a separate published system, and the dependency is genuine. No load-bearing circularity.\n\nBottom line: this is a solid, readable paper that should go to review. The reviewers should push on the depth-scale issue and demand error bars/protocol details. It's not a desk reject.\n\nRecommendation: send to peer review; I'd expect a conditional accept after the depth calibration is either fixed or honestly qualified.","headline":"Novel integration with a plausible but unverified metric-depth assumption; worth a careful referee, not a desk reject.","tokens_in":14964,"tokens_out":2764,"would_cite":true,"duration_ms":31976,"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":"A single RGB camera, no 3D map or trained controller, can drive a robot to object goals by fusing topological plans with terrain-aware metric control.","keywords":["topometric navigation","object-level topological map","traversability-aware control","BEV cost map","monocular depth","zero-shot navigation","CLIP traversability","instance image goal navigation"],"falsifier":"Run TANGO on the same HM3D episodes with simulator-provided depth in place of Depth-Anything while keeping every other component fixed. If the success rate rises by more than a few points (or if the Depth-Anything version exhibits collisions that the ground-truth-depth version avoids), then the metric controller's correctness depends on the monocular model's scale calibration. A complementary measurement: in real-world deployment, compare the planned BEV trajectory against a known obstacle layout and check whether planned paths ever clip walls or objects by more than the robot's radius.","tokens_in":14087,"feed_emoji":"🤖","tokens_out":7940,"duration_ms":73706,"temperature":0.7,"pith_summary":"This paper claims that a robot can perform long-horizon, open-set navigation toward object goals using a single RGB camera, with no 3D map, no learned controller, and no environment-specific training. The system couples an object-level topological global planner with a local metric planner: the global planner decides which visible object to approach based on graph connectivity to an (often distant) goal, and the local planner projects traversable image segments into a bird's-eye-view cost map using monocular depth, then plans a short collision-free path with Dijkstra's algorithm. In the zero-shot, no-ground-truth setting on the HM3D benchmark, the integrated system outperforms both a learned pixel-navigation controller and the topological controller on which it builds, across short, medium, and long trajectories, and the same modular pipeline is demonstrated on a real robot. The authors argue the method's practical benefit is deployability: it relies on foundation models that can be replaced as better ones appear, rather than on task-specific training.","feed_headline":"Robot navigates to object goals with one camera, no 3D map","feed_subtitle":"TANGO fuses object-level topological plans with terrain-aware metric control, beating learned baselines by up to 18 points.","key_machinery":"The central mechanism is the topometric controller: at every timestep it (1) segments the current RGB image into object masks, (2) labels each segment traversable or not with CLIP text queries (floor, ground, rug), (3) projects the traversable masks and the selected sub-goal segment into a metric bird's-eye-view (BEV) cost map using monocular depth from Depth-Anything, (4) selects the farthest projected point inside the lowest-cost sub-goal segment as the local target, and (5) plans a shortest path in the BEV cost map with Dijkstra's algorithm, driving a proportional yaw controller. This loop converts the topological planner's coarse 'head toward this object' signal into a continuous, obstac","core_discovery":"On the paper's own terms, the key result is that adding a traversability-aware metric controller on top of an object-level topological planner materially improves navigation success. In the no-ground-truth setting, success on easy (1-3m), hard (3-5m), and full (8-10m) trajectories is 61.76%, 43.14%, and 21.57%, versus 43.56%, 34.56%, and 13.73% for the topological-only baseline and 51.96%, 39.22%, and 14.0% for the learned pixel controller. The ablations attribute the larger remaining gap to perception—segment matching from the current view to the map—not to the metric controller: swapping in simulator ground-truth depth and traversability costs only about 5% on hard trajectories, while grou","pith_inferences":["If the monocular depth model's scale is inconsistent across viewpoints, the BEV cost map will be geometrically wrong; a direct test would be to substitute an onboard depth sensor or a SLAM-based scale correction while keeping everything else fixed and measure the resulting success and collision rates.","The ablation pattern (perception, not depth, drives the remaining 18% gap) suggests that future work should concentrate on open-set segment association—e.g., stronger visual-language matching—rather than on improving depth or traversability estimates.","The auto-switch mechanism hints at a broader principle: explicit uncertainty signals ('no traversable segment visible') can be used to trigger a fallback behavior, and this could be exported to other hybrid planners that blend learned and reactive control.","An immediate testable extension is to run the same pipeline on outdoor or off-road scenes with adapted text prompts (dirt, grass, gravel) to see whether the CLIP-based traversability filter keeps its open-set promise beyond indoor floors."],"forward_implications":["Robots could be deployed in new indoor environments using only a single teach drive (or any object-level map) and a monocular camera, with no simulation training or domain-specific fine-tuning.","Because the stack is modular, improving any component—segmentation, depth, text-based traversability, or segment matching—should directly improve navigation without redesigning the controller.","The 'seen but unvisited' result suggests object-level topological maps encode more than a route: the same graph can support goals the robot has never approached, which would extend teach-and-repeat systems into open-vocabulary tasks.","The auto-switch design offers a general pattern: letting a myopic but reliable controller take over when the learned traversability signal is missing is a cheap way to improve robustness in tight spaces."],"supporting_citations":[{"why":"Object-level topological map and global planner that produces sub-goal cost masks; TANGO replaces its controller with a metric one.","marker":"[8]"},{"why":"Monocular depth network used to project traversable segments and sub-goal points into the metric BEV cost map.","marker":"[14]"},{"why":"Segment Anything provides the open-set image segmentation underlying both object sub-goals and traversability masks.","marker":"[13]"},{"why":"CLIP text queries classify segments as traversable surfaces (floor, ground, rug).","marker":"[15]"},{"why":"Pixel-guided learned navigation baseline that TANGO is compared against on the HM3D benchmark.","marker":"[11]"},{"why":"Habitat-Matterport 3D dataset supplies the 36 simulated environments used for evaluation.","marker":"[79]"},{"why":"InstanceImageNav validation set defines the 108 episodes and the success metric.","marker":"[80]"},{"why":"LightGlue local feature matcher performs the segment-level data association between current view and map that the ablations identify as the main perception bottleneck.","marker":"[77]"}],"fun_headline_variants":["One camera, no map: TANGO routes to objects via topological+metric","Zero-shot object nav without 3D maps or trained policies","TANGO beats learned baselines by 18pts using terrain-aware control","Single RGB camera drives topological planning and traversability"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The system treats monocular depth from Depth-Anything as metric-scale depth without calibration, scale factor, or uncertainty model; if the network's relative depth cannot be converted to geometrically consistent metric coordinates across views, the BEV cost map and the selected sub-goal point will be wrong, breaking path planning and collision avoidance.","fun_headline_variants_meta":{"raw":{"variants":["One camera, no map: TANGO routes to objects via topological+metric","Zero-shot object nav without 3D maps or trained policies","TANGO beats learned baselines by 18pts using terrain-aware control","Single RGB camera drives topological planning and traversability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000228,"raw_usage":{"total_tokens":1328,"prompt_tokens":773,"completion_tokens":555,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":489}},"tokens_in":517,"tokens_out":555,"duration_ms":6510,"temperature":1.0,"reasoning_tokens":489,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T20:15:25.086925+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TANGO on the same HM3D episodes with simulator-provided depth in place of Depth-Anything while keeping every other component fixed. If the success rate rises by more than a few points (or if the Depth-Anything version exhibits collisions that the ground-truth-depth version avoids), then the metric controller's correctness depends on the monocular model's scale calibration. A complementary measurement: in real-world deployment, compare the planned BEV trajectory against a known obstacle layout and check whether planned paths ever clip walls or objects by more than the robot's radius.","supporting_citations":[],"review_version":1}