{"id":"177fb4f1-f720-4e1e-a3d0-572ec1a9ba08","arxiv_id":"2502.02664","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A dual-mode SDF trajectory optimizer, using object and robot body neural SDFs plus memory, is reported to achieve 98% success in simulated dynamic indoor navigation.","lead":"This paper presents a robot navigation method that builds signed distance fields from RGB-D camera data and combines two different SDF-based trajectory optimizers to avoid moving obstacles indoors. The authors report a 98% success rate in simulation, although the exact figure is inconsistent in the text, and they demonstrate the approach on a real Turtlebot4.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Runtime selection of per-instance DeepSDF latent codes is never specified; if the iGibson 80-object catalog supplies them, the 98% result tests closed-set retrieval, not compositional inference, and the real-world claim is unquantified.","rationale":"The reader correctly flags the quantitative inconsistency and the 80-object closed set. My stress-test targets the mechanism underneath that closed-set concern: even within the iGibson benchmark, the paper never says how the per-instance latent code z_i is selected. Without that, the Scene SDF pipeline is either a catalog lookup or an unimplemented online inference. Both readings weaken the abstract's general claim. A concrete code audit and a held-out-instance rerun would settle which reading is correct. This does not change my overall confidence — the method is plausible and the modular baseline is meaningful — so the conditional verdict stands. In addition, the 98.0 vs 95.67 discrepancy in §IV.C should be corrected as part of the revision, but it is secondary to the specification gap.","tokens_in":11151,"tokens_out":7440,"duration_ms":74747,"concrete_test":"With the released code, trace the PCAlign path in Algorithm 1 and identify the source of z_i for a detected object (e.g., a hard-coded catalog lookup, a class-level canonical code, or an online auto-decoder fit). Then rerun the 300-simulation benchmark with a held-out set of object instances whose shapes are absent from the DeepSDF training set and the internal catalog. If the success rate drops materially, the headline 98% depends on closed-set priors; if it holds, the generalization concern is substantially answered.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central Scene SDF stage is not fully specified. Section III.B defines each object SDF as Ω_{o_i}(·, z_i) with a per-shape DeepSDF latent code, and Section III.C's alignment loss (Eq. 1) requires evaluating Ω_{o_i} at transformed points. Yet Algorithm 1 receives only a YOLOv5 class label and never explains how z_i is obtained. If z_i is taken from a stored table of the 80 iGibson objects, the dual-mode success rate is a closed-set benchmark by construction and tells us nothing about novel object instances; if z_i is inferred online from the partial point cloud, that inference procedure — its objective, initialization, and cost — is entirely absent, and Eq. (1) is not computable as written. The brief real-world section sidesteps this by replacing DeepSDF with convex-hull SDFs, so it does not validate the compositional mechanism. Because the paper's central contribution is the composition of learned object priors, this omission is load-bearing for the claimed general dynamic-indoor navigation, independent of the 98.0/95.67 success-rate discrepancy.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a dual-mode trajectory optimization framework for mobile robot navigation in dynamic indoor scenes using only an onboard RGB-D camera. The primary mode queries a learned SDF of the robot body, swept along the trajectory, at obstacle point-cloud points; the secondary mode aligns per-object DeepSDF latent-shape priors to detected objects, composes them with an iSDF static-scene SDF, and optimizes the trajectory against the composite field. The two modes are combined by triggering the scene-SDF stage only when the robot-SDF stage returns a colliding trajectory. The method is evaluated in iGibson 2.0 on 300 simulated runs across 10 scenarios and additionally demonstrated on a Turtlebot4 in a real office setting. The headline claim is a 98% success rate, 14.4% higher than baseline, with comparable amortized planning time.","tokens_in":11401,"tokens_out":4846,"duration_ms":49918,"significance":"If validated, the compositional approach is a useful step toward practical neural-SDF navigation: it avoids per-timestep retraining of a scene SDF, operates under partial observability, and is modular, allowing individual SDF components to be swapped. The authors provide code and video links, which is a strength for reproducibility. However, the central quantitative claim is currently undercut by internal inconsistencies in the reported success rate and by an unspecified mechanism for obtaining the per-instance DeepSDF latent codes that Eq. (1) requires. The real-world demonstration replaces DeepSDF with convex-hull SDFs, so it does not validate the compositional learned-prior mechanism. These issues need to be resolved before the paper's generalization claims can be accepted.","major_comments":[{"comment":"The headline success rate is reported inconsistently: the abstract states 98%, Section IV.C states 95.67%, and Table I reports 98.0%. Since the abstract's 14.4% improvement and the entire quantitative contribution rest on this number, the authors must identify the correct value and explain the discrepancy. In addition, with 300 simulations, the binomial standard error is roughly 1–2 percentage points, so differences such as 98.0% versus 96.33% are not obviously significant; confidence intervals or significance tests should be reported.","section":"Abstract / Section IV.C / Table I"},{"comment":"The source of the per-instance DeepSDF latent code z_i is never specified. Algorithm 1 receives only a YOLOv5 class label and the point cloud, but Eq. (1) evaluates Ω_{o_i}(R_i x̄) with a latent code z_i. If z_i is retrieved from a table of the 80 iGibson objects, the experiment is closed-set by construction and does not test compositional generalization to novel instances; if z_i is inferred online from the point cloud, the inference objective, initialization, and cost are absent, making Eq. (1) not computable as written. This is load-bearing for the claimed general dynamic-indoor navigation and must be clarified.","section":"Algorithm 1 / Section III.C, Eq. (1)"},{"comment":"The composition Ω(x) = min(Ω_s(x), min_i Ω_{o_i}(H_i x)) does not define a signed distance field in the workspace unless each H_i is an isometry. The transformation H_i includes a scaling S_i, and the text specifies s = [1,1,1,γ], so the object-SDF values evaluated at H_i x are not Euclidean signed distances in the workspace. The zero level set and sign information remain meaningful, but the claimed 'SDF of the visible scene' and the gradient interpretation used in Eq. (3) are not mathematically justified as stated. The authors should either restrict the alignment to isometries, account for the non-isometric Jacobian, or explicitly present the composition as a pseudo-SDF whose gradients are only approximate.","section":"Section III.C, Eq. (2)"},{"comment":"The '14.4% higher than baseline' claim does not name the baseline. In Table I, Dual Mode has 98.0% and DWA+PC-Mem has 85.67%; the relative improvement (98.0−85.67)/85.67 is approximately 14.4%, suggesting DWA+PC-Mem is the intended baseline. The baseline should be named in the abstract and in Section IV.C, and the comparison should be framed carefully, since the margin over the Robot SDF pipeline (96.33%) is only 1.67 percentage points.","section":"Abstract / Section IV.B / Table I"},{"comment":"The real-world evaluation is anecdotal and does not exercise the compositional Scene SDF pipeline: the convex-hull SDFs replace the DeepSDF priors, so the real-world results do not validate the central learned-composition mechanism. The section should report the number of runs, success/failure outcomes, and any failure modes, and should be positioned as a demonstration of the Robot SDF pipeline (or as a separate convex-hull variant) rather than as evidence for the full dual-mode method.","section":"Section IV.E"}],"minor_comments":[{"comment":"The bracket parameters δ_l and δ_u are said to be 'progressively reduced,' but no schedule, stopping criterion, or final values are given; please specify the update rule and convergence condition.","section":"Section III.C, Eq. (1)"},{"comment":"The indicator α_i is written as I(Ω_r(M_i x) < ζ) inside the sum over x_o, but the subscript i refers to trajectory points while x is an obstacle point; clarify whether the indicator is evaluated per obstacle point, per trajectory point, or as a joint condition.","section":"Section III.D, Eq. (4)"},{"comment":"The text says the Scene SDF pipeline 'requires a minimum distance assumption,' but this assumption is never defined; please state the minimum distance and its role in the algorithm.","section":"Section III.F"},{"comment":"The scalability plot reports trends without error bars or per-density sample sizes; please add error bars or confidence intervals and state how many simulations were run per obstacle-density level.","section":"Section IV.D / Figure 6"},{"comment":"There are several language issues, including 'robot-centeric' in the Introduction and the repeated use of 'formidable workspace' where 'occupied workspace' or 'forbidden workspace' is likely intended; a careful proofread is recommended.","section":"Introduction and throughout"},{"comment":"The hardware and simulator setup are clear, but the number of YOLOv5 detections that were incorrect or missing in the simulation is not reported; this matters because the Scene SDF pipeline depends on detection outputs.","section":"Section IV.A"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the modular design is attractive, but the paper currently has an unresolved internal inconsistency in the headline result and a missing specification of the latent-code mechanism that the main method depends on. I would ask for a revised version that states the success rate unambiguously, reports statistical uncertainty, names the baseline, and either specifies the latent-code inference procedure or explicitly reframes the experiments as closed-set retrieval. The mathematical issue with the non-isometric composition in Eq. (2) also needs a clear response, since it affects the paper's central formal claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, the short version: this is a credible engineering paper that deserves referee time, but the numbers need a consistency pass and the compositional part has a specification gap you should push on.\n\nThe genuinely new bit is the dual-mode operation: a fast robot-body SDF query runs most of the time, and a slower scene-composition SDF step is triggered only when the fast path gets stuck. That is a sensible way to amortize cost in dynamic scenes, and the memory modules for out-of-view points are a practical touch. The iGibson evaluation is serious: 300 runs per method, three external baselines plus two ablations, and a clutter-scalability study. The method section is mostly readable and the equations hang together.\n\nSoft spots, in order of weight. First, the headline number is inconsistent: the abstract and Table I say 98.0%, Section IV.C says 95.67%. That is the kind of thing that makes referees wonder what else was cleaned up. Second, the '14.4% higher than baseline' improvement is not tied to a named baseline. Third, no confidence intervals on any success rate. Fourth, and most important, the Scene SDF pipeline never says how the per-instance DeepSDF latent code z_i is obtained. Algorithm 1 gets only a YOLO class label, then calls PCAlign with Omega_{o_i}. If z_i is pulled from a table of the 80 iGibson objects, that is a closed-set retrieval benchmark, not compositional inference on novel shapes, and the paper should say so. If it's inferred online, the inference procedure is missing and Eq. 1 is not computable as written. The real-world section dodges this by replacing DeepSDF with convex-hull SDFs, so it does not validate the compositional claim either.\n\nNone of this kills the core idea. The system-level contribution is still real and the benchmark is meaningful. But the paper currently overstates what is demonstrated, and the missing z_i specification is a reproducibility blocker.\n\nWho is this for? People building reactive local planners for indoor robots, and anyone working on SDF-based scene representations for navigation. It's a useful read, but I would not treat the 98% as durable until the numbers are reconciled and the latent-code question is answered.\n\nRecommendation: send it to review. The issues are addressable with a revision; the system deserves a serious referee's time.","headline":"A practical dual-mode SDF navigation system with real promise, but the per-object latent code selection is underspecified and the headline success rate is inconsistent.","tokens_in":11962,"tokens_out":2702,"would_cite":true,"duration_ms":25282,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A dual-mode composite neural SDF framework lets a mobile robot with an onboard RGB-D camera plan collision-free trajectories through dynamic indoor scenes without retraining the environment representation.","keywords":["neural signed distance fields","robot navigation","dynamic environments","trajectory optimization","RGB-D perception","object pose alignment","dual-mode planning","iGibson benchmark"],"falsifier":"Re-run the iGibson 2.0 benchmark with obstacles drawn from categories not seen during training, or with objects placed in tipped or rotated poses, and compare success rates; if the dual-mode success rate falls toward the EgoTrajOpt baseline (78.33%) or the scene-SDF stage fails to align, the generalization claim fails. Additionally, reconcile the reported success rate: Section IV.C states 95.67% while the abstract and Table I state 98.0% for the same 300 simulations, so a re-run should establish the actual number.","tokens_in":10922,"feed_emoji":"🤖","tokens_out":7615,"duration_ms":70320,"temperature":0.7,"pith_summary":"The paper argues that a robot can navigate a changing indoor environment with only a single onboard RGB-D camera by composing pre-trained neural signed distance fields (SDFs) for known objects with a static scene SDF, rather than rebuilding the environment model at each step. It proposes a dual-mode planner: a fast primary stage that queries the robot body's SDF against the visible obstacle point cloud, and a slower secondary stage that aligns each detected object's point cloud to its shape prior and composes the resulting SDFs to supply accurate collision gradients. The secondary stage is invoked only when the primary stage gets trapped in a local minimum, so the average planning time stays low. If the claim holds, robots can adapt to moved chairs, toys, and other everyday objects from local perception alone, without external tracking or full-scene retraining.","feed_headline":"Dual-mode neural SDF navigation hits 98% in dynamic rooms","feed_subtitle":"A fast robot-body SDF stage handles most cases; a scene-SDF stage rescues it from local minima without retraining per step.","key_machinery":"The load-bearing mechanism is the composite signed distance field $\\Omega(x) = \\min(\\Omega_s(x),\\, \\min_i \\Omega_{o_i}(H_i x))$ over the visible scene, built by aligning each detected object's point cloud to its DeepSDF shape prior through objective (1), which progressively narrows the allowed signed-distance band around the zero level set. Because both the object priors and the static scene field are neural, $\\Omega$ is differentiable and supplies both collision penalties and push-out gradients to a trajectory optimizer minimizing the cost in (3). The fast pipeline replaces this with the robot-body SDF $\\Omega_r$ queried at obstacle points via (4), so each mode is a different way of evaluating collision information, and the dual-mode trigger decides which one runs.","core_discovery":"The paper's central claim is that a dynamic indoor scene can be navigated from a single onboard RGB-D camera by composing pre-trained neural SDFs instead of retraining an environment SDF each time the scene changes. The full scene SDF is formed as $\\Omega(x) = \\min(\\Omega_s(x),\\, \\min_i \\Omega_{o_i}(H_i x))$, where $\\Omega_s$ is the static background field from a scene-level network, each $\\Omega_{o_i}$ is an object-level shape prior, and $H_i$ maps workspace points into that prior's normalized domain. A fast planning stage instead queries the robot body's own SDF at the obstacle point cloud, yielding collision costs with well-defined gradients; only when that stage remains in collision does the framework compute the composite scene SDF and re-optimize from its gradients. On 300 randomized iGibson 2.0 simulations with 80 household objects, the dual-mode pipeline claims a 98.0% success rate (Section IV.C reports 95.67%) with 15.73 s average plan time, beating the EgoTrajOpt baseline's 78.33% at comparable cost.","pith_inferences":["An obvious extension is to replace the fixed detector and shape priors with category-agnostic instance segmentation and a generative shape model, which would relax the known-object requirement and let the planner handle unseen household items.","The in-plane-only pose alignment is the likely failure point in the real world: tipping an object beyond a few degrees would make the point cloud fail to match the prior's zero set, so a benchmark with arbitrary orientations would quantify how much of the reported success depends on upright objects.","The same composition rule could be applied to a manipulator arm by substituting the arm's SDF for the robot body's SDF, turning the framework into a whole-body motion planner for dynamic tabletop scenes.","The reported success-rate inconsistency (95.67% in Section IV.C versus 98.0% in Table I and the abstract) should be resolved by re-running the 300-simulation benchmark, since the absolute number matters for downstream comparisons."],"forward_implications":["A robot using this planner can react to objects moved into its path mid-run, because the scene representation is recomposed from the current RGB-D frame rather than retrained.","Most timesteps use the cheap robot-body stage, so amortized planning stays near the 13.87 s of the fast stage rather than the 55.97 s of the full scene-SDF stage.","Failures of the fast stage are not terminal: the scene-SDF stage's better gradients turn a local-minimum trajectory into a collision-free initialization that the fast stage can continue refining.","The composite field's differentiability means the same trajectory optimizer can switch between modes without changing the cost structure, only the source of the SDF queries.","The paper's obstacle-density analysis indicates the dual-mode success rate degrades more slowly than the baselines as clutter increases."],"supporting_citations":[{"why":"Supplies the object-level DeepSDF priors that the scene-SDF stage aligns point clouds to.","marker":"[4]"},{"why":"Supplies the scene-level iSDF from posed depth frames, used as the static background field and to filter object points.","marker":"[5]"},{"why":"Supplies the YOLOv5 detections that pair bounding boxes with object point clouds.","marker":"[36]"},{"why":"Provides the EGO-Planner baseline (EgoTrajOpt) whose success rate the dual mode must beat.","marker":"[29]"},{"why":"Provides the iGibson 2.0 simulator with 80 household objects used for all benchmark comparisons.","marker":"[34]"},{"why":"Provides the paradigm of querying a continuous implicit robot-body SDF at obstacle point clouds, which the fast stage uses.","marker":"[33]"},{"why":"Shows predicted composite signed-distance fields for dynamic motion planning, the prior work this framework extends to onboard, partial-observability navigation.","marker":"[1]"}],"fun_headline_variants":["Dual-mode neural SDFs navigate dynamic rooms at 98% success","Composed neural SDFs avoid retraining, hit 98% in dynamic rooms","No-retrain neural SDF composition hits 98% in dynamic rooms","Single RGB-D camera, composed SDFs, 98% nav success","Neural SDFs composable on the fly for 98% dynamic navigation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that every movable obstacle belongs to a known category with a pre-trained shape model, can be spotted by an RGB object detector, and is aligned to its point cloud by moving and rotating around a single vertical axis—assumptions that hold by construction in the simulator, since all obstacles come from an 80-object set.","fun_headline_variants_meta":{"raw":{"variants":["Dual-mode neural SDFs navigate dynamic rooms at 98% success","Composed neural SDFs avoid retraining, hit 98% in dynamic rooms","No-retrain neural SDF composition hits 98% in dynamic rooms","Single RGB-D camera, composed SDFs, 98% nav success","Neural SDFs composable on the fly for 98% dynamic navigation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000822,"raw_usage":{"total_tokens":3627,"prompt_tokens":1008,"completion_tokens":2619,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":624,"completion_tokens_details":{"reasoning_tokens":2516}},"tokens_in":624,"tokens_out":2619,"duration_ms":18498,"temperature":1.0,"reasoning_tokens":2516,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T11:32:04.436608+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the iGibson 2.0 benchmark with obstacles drawn from categories not seen during training, or with objects placed in tipped or rotated poses, and compare success rates; if the dual-mode success rate falls toward the EgoTrajOpt baseline (78.33%) or the scene-SDF stage fails to align, the generalization claim fails. Additionally, reconcile the reported success rate: Section IV.C states 95.67% while the abstract and Table I state 98.0% for the same 300 simulations, so a re-run should establish the actual number.","supporting_citations":[{"cited_title":"isdf: Real-time neural signed distance fields for robot perception,","cited_arxiv_id":null,"evidence_quote":"Supplies the scene-level iSDF from posed depth frames, used as the static background field and to filter object points."},{"cited_title":"Yolov5 by ultralytics,","cited_arxiv_id":null,"evidence_quote":"Supplies the YOLOv5 detections that pair bounding boxes with object point clouds."},{"cited_title":"igibson 2.0: Object-centric simulation for robot learning of everyday household tasks,","cited_arxiv_id":null,"evidence_quote":"Provides the iGibson 2.0 simulator with 80 household objects used for all benchmark comparisons."},{"cited_title":"Continuous implicit sdf based any-shape robot trajectory optimization,","cited_arxiv_id":null,"evidence_quote":"Provides the paradigm of querying a continuous implicit robot-body SDF at obstacle point clouds, which the fast stage uses."},{"cited_title":"Predicted composite signed- distance fields for real-time motion planning in dynamic environ- ments,","cited_arxiv_id":null,"evidence_quote":"Shows predicted composite signed-distance fields for dynamic motion planning, the prior work this framework extends to onboard, partial-observability navigation."}],"review_version":1}