{"id":"1cfd89d4-f89f-4f6d-bb3e-f86119a2e881","arxiv_id":"2411.15915","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A multi-robot scan-n-print framework for wire arc additive manufacturing uses laser-scanned layer heights and adjusted torch speed to improve geometric fidelity over open-loop printing.","lead":"This paper presents a three-robot system that scans each printed metal layer with a laser and adjusts the next layer's welding speed to keep the part height uniform. On test prints in aluminum and steel, the closed-loop method reduced layer height variation compared with open-loop printing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Closed-loop scans add inter-layer dwell time absent from the open-loop baseline; replay data suggest the edge-excluded height improvement may be a thermal-history confound rather than the speed-control law.","rationale":"The central claim is that the closed-loop speed adjustment improves layer-height uniformity by 66% (wall) and 53% (blade) over open-loop printing. For that claim to hold, the observed improvement must be attributable to the control law. The paper's step-wise implementation inserts a full laser-scanning pass between layers; the open-loop baseline has no such pause. Aluminum WAAM is known to be sensitive to inter-layer cooling, so the dwell time alone could reduce height variation. Evidence within the paper strengthens this concern: in the repeatability test (Sec. 4.6), replaying the closed-loop speed profile in an open-loop run (without scanning pauses) reproduces the full-piece quality, but the text explicitly states that with edges excluded the replayed prints perform similarly to the open-loop baseline. Since Table 2 claims a 60% edge-excluded improvement for the closed-loop run, the speed profile itself does not carry the interior correction. The remaining confounds are the online feedback and the inter-layer dwell time. The paper provides no matched-dwell open-loop control, and the continuous-mode cylinder result (Sec. 4.5) has no open-loop baseline, so the confound is not ruled out. This is more load-bearing than the reader's weakest assumption (model validity): even a perfect model would not rescue the central claim if the improvement is an artifact of the experimental protocol. I therefore propose a dwell-matched open-loop control experiment as the decisive check. If the confound is confirmed, the paper's causal claim is rejected; if not, the existing CONDITIONAL verdict stands.","tokens_in":13097,"tokens_out":17011,"duration_ms":159627,"concrete_test":"Print one additional open-loop wall (100 ipm, 5 mm/s constant speed) with exactly the same inter-layer dwell time as the step-wise scan-n-print run (i.e., pause for the equivalent scan duration after each layer, do not modify speeds), and compare its edge-excluded layer-height STD to Table 2 (baseline 0.33 mm, correction 0.13 mm). If the dwell-matched open-loop STD drops to near 0.13 mm, the claimed 60% without-edge improvement is a thermal-history artifact; if it remains near 0.33 mm, the speed-control law is responsible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The step-wise closed-loop experiments (Sections 4.2-4.4) insert a full laser-scanning pass between every layer, whereas the open-loop baselines are printed without these pauses. In WAAM, especially for aluminum (ER 4043), inter-layer dwell time changes thermal history and can by itself reduce layer-height variation. Section 4.6's replay test isolates the speed-profile contribution: when the recorded speed commands are replayed in an efficient open-loop run (no scanning), the full-piece quality matches the closed-loop part, but the text states that with edge regions excluded 'the replayed prints perform similarly as the open loop.' Yet Table 2 reports a 60% edge-excluded height-STD improvement (0.13 mm vs 0.33 mm) for the closed-loop wall. If the recorded speed profile were responsible for that interior improvement, replaying it should reproduce it; it does not. The only remaining differences between the closed-loop run and the replay are the online feedback and the inter-layer dwell time. This leaves open the possibility that the reported without-edge improvement (and a substantial part of the 66% headline figure) is due to the cooling pause rather than to the model-based speed adjustment. The paper does not report inter-layer timing for the baseline versus the scan-n-print runs, so the confound is unaddressed.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a multi-robot scan-n-print framework for Wire Arc Additive Manufacturing (WAAM). A sensing robot with a laser line scanner measures the deposited layer height after each weld layer, and an identified log-log model relating torch speed to deposition height is inverted to compute speed commands for the next layer, aiming to reduce layer-height variation. The framework is demonstrated on an aluminum wall, a fan-blade-like geometry, a steel wall, and a cylinder, with reported improvements in layer-height standard deviation (e.g., 66% for the aluminum wall) and CAD-comparison errors. A repeatability test replays the recorded speed profiles in open-loop printing to assess whether the learned profiles transfer to subsequent parts.","tokens_in":13382,"tokens_out":3776,"duration_ms":32958,"significance":"If the reported improvements are robust, the paper makes a useful engineering contribution: it integrates multi-robot coordination, laser scanning, model identification, and closed-loop speed control in a real WAAM testbed, and it shares source code, datasets, and a video. The model identification is reasonable, with slopes close to the theoretically expected -0.5, and the application to two materials and several geometries adds breadth. However, the repeatability test in Section 4.6 raises a serious confound: the closed-loop runs include inter-layer scanning pauses that are absent from the open-loop baselines, and the replayed speed profiles do not reproduce the interior improvement. Because the paper's central claim is the empirical superiority of the closed-loop approach, this confound must be addressed before the improvement can be attributed to the control law rather than to thermal-history effects. The lack of error bars and the post hoc edge-exclusion threshold further weaken the quantitative claims. The manuscript is likely to be acceptable after substantial revision and additional experiments or analysis.","major_comments":[{"comment":"The repeatability test undermines the attribution of the without-edge improvement to the speed-control law. The text states that when edge regions are excluded, \"the replayed prints perform similarly as the open loop,\" yet Table 2 reports a 60% without-edge improvement (0.13 mm vs 0.33 mm) for the closed-loop wall. Since replaying the recorded speed commands in an efficient open-loop run (without scanning) does not reproduce the interior improvement, the only remaining differences between the closed-loop run and the replay are online feedback and the inter-layer dwell time introduced by the scanning pass. The paper does not report inter-layer timing for the baseline versus the scan-n-print runs, so the possibility that the without-edge improvement is largely a thermal-history effect of the cooling pause is left unaddressed. This directly affects the headline 66% full-piece improvement, whose without-edge component is the more controlled measure of process improvement. The authors should add an open-loop baseline that includes the same inter-layer pauses (or a scan-only pass without speed adjustment) to isolate the controller's contribution, or otherwise quantify the thermal-history effect.","section":"Section 4.6, Table 2, Figure 19(b)"},{"comment":"The edge-exclusion threshold of 7.5 mm from the arc on/off points is introduced post hoc and its sensitivity is not examined. The reported without-edge improvements rest entirely on this threshold: for aluminum the without-edge improvement is 60% (0.33 mm vs 0.13 mm), but for steel it is only 10% (0.31 mm vs 0.28 mm), suggesting a strong material and threshold dependence. Without a sensitivity analysis over a range of exclusion widths (e.g., 5, 7.5, 10, 12.5 mm) or a physical justification of the threshold, the without-edge metric is not robust evidence for interior uniformity improvement. The authors should either justify the threshold from arc on/off physics or show that the conclusion is insensitive to the choice.","section":"Section 4.2, Table 2, Section 4.4, Table 6"},{"comment":"All improvement percentages are based on single runs per condition, with no error bars, confidence intervals, or run-to-run variability reported. Given the inherent noise in WAAM and the visible variability between Repeat 1 and Repeat 2 in the repeatability test, the reader cannot assess whether the 66% and 53% improvements are statistically meaningful or within process noise. The paper should report the per-layer standard deviation distributions (e.g., box plots) and, ideally, multiple baseline and closed-loop runs, or at least a sensitivity analysis of the computed metrics to measurement noise. This is load-bearing because the central claim is empirical improvement, and the current tables present only point estimates.","section":"Section 4.2-4.4, Tables 2, 4, 5, 6"},{"comment":"The identified model ln(Δh) = a ln(v) + b is fitted from a dedicated calibration schedule with speeds decreasing from 20 mm/s to 2 mm/s, and the model is assumed to remain valid for all subsequent layers and geometries in the closed-loop experiments. The closed-loop runs at 100 ipm use speeds within the calibration range, but the blade's short end segments and the edge regions where arc on/off effects dominate may violate the model's monotonicity assumption, as the paper itself notes that defects concentrate at the edges. The authors do not validate the model's predictive accuracy on the actual closed-loop runs (e.g., comparing the Δh predicted from the commanded speeds with the measured layer heights). Adding such residual analysis would strengthen the claim that Eq. (1) produces the intended height corrections and would help rule out model mismatch as an alternative explanation for the observed improvements.","section":"Section 3, Eq. (5), Section 4.2"}],"minor_comments":[{"comment":"There are several typos and grammatical errors, including \"addition 6-dof robot\" (Section 1, should be \"additional\"), \"measurements the layer height\" (Section 2.2, should be \"measures\"), \"an average smoother\" (Section 2.3, should be \"an averaging smoother\"), \"The proposed techniques works well\" (Section 1, should be \"work well\"), and \"the close-loop approach\" (Section 4.4, should be \"closed-loop\").","section":"General"},{"comment":"The layer-height STD plots do not have labeled axes or explicit legends identifying which curve corresponds to the baseline and which to the correction. Adding axis labels and a legend would make the figures self-contained.","section":"Figures 10 and 16"},{"comment":"The table reports RMSE values in mm, but the fitted model is in log space (ln Δh vs ln v). Please clarify whether the RMSE is computed in the original height units after back-transformation or in log units, and specify the calculation.","section":"Table 1"},{"comment":"The text says \"the uniformity between the open-loop and the closed-loop approach is comparable\" but then claims a trend of error accumulation in the open-loop case. This is somewhat contradictory; please clarify whether the without-edge comparison is comparable overall or shows a layer-dependent trend.","section":"Section 4.4, Figure 16(b)"},{"comment":"The continuous scan-n-print of the cylinder is presented without an open-loop baseline for the same geometry, so the reported reduction in standard deviation is only an internal comparison across layers. State the baseline explicitly or note that the baseline is not included.","section":"Section 4.5"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely problem in robotic WAAM and provides a useful open-source implementation. The main concern is the thermal-history confound identified in Section 4.6; the authors' own replay data suggest that the interior improvement may not be due to the speed-control law. This is fixable with additional experiments (e.g., an open-loop baseline with identical inter-layer dwell times) or a careful analysis of the existing data. The post hoc edge threshold and missing error bars are secondary but should be addressed in revision. The self-citations to the group's prior architecture are appropriate and not excessive. The manuscript fits the journal's scope, but the central empirical claim needs stronger support before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper delivers a genuine systems integration: three coordinated robots, laser scanning, layer-height feedback, and a continuous look-ahead mode, with public code and data. That is worth something. Second, the central quantitative claim needs a close look. The closed-loop runs insert a scanning pass between every layer; the open-loop baseline has no such pause. In aluminum WAAM, inter-layer cooling alone can reduce height variation. The paper's Section 4.6 replay test is the natural control: they replay the recorded speed profiles in an efficient open-loop run, no scanning, and the full-piece quality matches. But for edge-excluded regions, the replayed prints \"perform similarly as the open loop\" — so the recorded speed commands do not reproduce the 60% interior improvement reported in Table 2. The paper does not report inter-layer timing for baseline versus scan-n-print, so the dwell-time confound is unaddressed. This is a serious soft spot because it directly affects the headline success metric.\n\nWhat is new: the three-robot integration, the continuous look-ahead mode at 100 Hz, and the repeatability transfer concept — learned speed profiles reusable in open loop. The deposition model is identified from [22], and the fitted slopes are close to -0.5, consistent with the derivation. The experiments cover aluminum and steel, wall and blade. The edge-defect correction is convincing and reproducible; the replay runs do reproduce the edge improvement, and that alone is practically useful.\n\nOther soft spots: no error bars or replicates on the reported percentages; the edge-exclusion threshold (7.5 mm) is post hoc and not varied; the CAD-level accuracy gains are modest (13% average for wall); the model's validity across thermal history is assumed; the continuous cylinder result has no open-loop baseline. The paper is honest in Section 4.6 — it explicitly says the correction is mostly for the edge — but that honesty undercuts the earlier overclaim in 4.2.\n\nOverall: worth a serious referee. It is a solid engineering demonstration, though a revision should report inter-layer timing, add replicates, and recalibrate the interior improvement claim. I would cite the system architecture and the replay result, not the 60% interior number.","headline":"Solid multi-robot WAAM integration with a real edge-correction result, but the paper's own replay test undermines its interior height-improvement claim by leaving dwell time as the likely confound.","tokens_in":13900,"tokens_out":2145,"would_cite":true,"duration_ms":20332,"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":"Closed-loop scanning cuts 3D-printed metal height error by 66%","keywords":["WAAM","wire arc additive manufacturing","closed-loop control","laser scanning","layer height regulation","torch speed control","multi-robot system","scan-n-print"],"falsifier":"Print a tall wall with more than 40 layers in scan-n-print mode and plot the per-layer height standard deviation; if the standard deviation trends upward with layer count or matches the open-loop baseline for layers beyond the calibration range, the model's cross-layer validity is falsified.","tokens_in":12910,"feed_emoji":"🤖","tokens_out":5668,"duration_ms":47394,"temperature":0.7,"pith_summary":"Wire arc additive manufacturing (WAAM) builds metal parts layer by layer, but the height of each layer varies and errors accumulate, especially with aluminum. This paper proposes a closed-loop 'scan-n-print' system in which a laser scanner measures each printed layer and the welding robot's torch speed is adjusted for the next layer based on an identified power-law model. On their testbed, the mean layer height standard deviation improved 66% for an aluminum wall and 53% for a turbine-blade-shaped part, and the corrected speed profile could be replayed to print a second part with similar quality.","feed_headline":"Closed-loop scanning cuts 3D-printed metal height error by 66%","feed_subtitle":"Laser-scanned layer heights adjust torch speed, and the learned profile can be replayed on later parts.","key_machinery":"The load-bearing element is the deposition-height model $\\ln(\\Delta h)=a\\ln(v)+b$, fitted from a calibration schedule of two base layers followed by two layers at each of several torch speeds from 20 mm/s down to 2 mm/s. Because the model is monotonic in speed, it can be inverted: for each motion segment of the next layer, the measured average height is compared with the target, and Eq. (1) gives the required torch speed. The control loop is closed by the scanning robot, whose wrist-mounted laser line scanner captures the top surface after each layer; point clouds are registered to the positioner frame using calibrated kinematics, cleaned, and sampled along the welding path to produce the height profile. Smoothing and speed caps are applied to avoid excessive accelerations, and leftover error is corrected in later layers.","core_discovery":"The central claim is that geometric precision in WAAM can be achieved by feedback control of torch travel speed alone, without altering wire feed rate or other welding parameters. Using a three-robot testbed (welding robot, positioner, and scanning robot with a laser line scanner), the authors show that the layer height profile can be measured after each deposition and that inverting an identified model $\\ln(\\Delta h)=a\\ln(v)+b$ yields torch speeds that keep subsequent layers close to the desired height. The model is derived from a mass-conservation argument that predicts a slope near $-0.5$, and the fitted slopes in Table 1 are close to this value. In experiments, the closed-loop approach reduces mean layer height standard deviation from 1.38 mm to 0.47 mm (66%) on a wall and from 0.85 mm to 0.24 mm (53%) on a blade, and also improves tracking of the CAD geometry.","pith_inferences":["One could extend the controller to also modulate wire feed rate, as the authors list as future work; jointly controlling both would decouple height and bead width, potentially giving finer control than speed alone.","The strong residual errors at the edges where the arc turns on and off suggest an explicit edge-compensation strategy (e.g., a precomputed speed pulse at segment boundaries) might yield further gains beyond the current layer-by-layer correction.","The replay result implies a teach-and-repeat manufacturing model: the first part pays the scanning cost, and subsequent parts reap the benefit, which is economically attractive for small-batch production."],"forward_implications":["If closed-loop speed control holds up, WAAM parts can be printed to tighter geometric tolerances without changing material feed parameters, reducing the need for post-process machining.","Because the learned speed profile can be replayed for subsequent copies of the same part, the scanning step can be dropped for later units, saving cycle time while retaining the corrected geometry.","The same framework transfers to a steel alloy and to a continuous look-ahead mode with a cooperating scanner, suggesting it is not restricted to one material or one print strategy.","The physical derivation of the model (slope near $-0.5$) hints that the speed-to-height relation may be transferable to other wire-fed materials with modest re-identification."],"supporting_citations":[{"why":"Supplies the deposition equation (mass conservation) from which the power-law model is derived.","marker":"[22]"},{"why":"Provides the multi-robot motion planning and software architecture that coordinates the three robots.","marker":"[3]"},{"why":"Hand-eye calibration for the 2D laser profile scanner, enabling point cloud registration.","marker":"[21]"},{"why":"ICP registration used to compare printed part to CAD model for accuracy evaluation.","marker":"[23]"},{"why":"Plug-and-play software architecture for coordinating multiple industrial robots and sensors, the control backbone.","marker":"[19]"}],"fun_headline_variants":["Torch speed feedback cuts WAAM height error by 66%","Laser-scan robot feedback halves metal 3D print error","Closed-loop speed control improves wire-arc printing accuracy","Multi-robot scan-n-print reduces metal height error by two-thirds","Feedback-controlled torch speed trims 3D-printed metal error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The identified model $\\ln(\\Delta h)=a\\ln(v)+b$, calibrated on two-layer runs at each speed, is assumed to hold for every later layer and for the short end segments where arc on/off effects dominate; if the model drifts with thermal history or geometry, the inverted speed commands will not produce the intended height fix.","fun_headline_variants_meta":{"raw":{"variants":["Torch speed feedback cuts WAAM height error by 66%","Laser-scan robot feedback halves metal 3D print error","Closed-loop speed control improves wire-arc printing accuracy","Multi-robot scan-n-print reduces metal height error by two-thirds","Feedback-controlled torch speed trims 3D-printed metal error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000265,"raw_usage":{"total_tokens":1634,"prompt_tokens":996,"completion_tokens":638,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":549}},"tokens_in":612,"tokens_out":638,"duration_ms":6246,"temperature":1.0,"reasoning_tokens":549,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:43:40.350000+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Print a tall wall with more than 40 layers in scan-n-print mode and plot the per-layer height standard deviation; if the standard deviation trends upward with layer count or matches the open-loop baseline for layers beyond the calibration range, the model's cross-layer validity is falsified.","supporting_citations":[{"cited_title":"The robotic easy teaching system incomputeraidedwelding,","cited_arxiv_id":null,"evidence_quote":"Supplies the deposition equation (mass conservation) from which the power-law model is derived."},{"cited_title":"Open-Source Software Architecture for Multi-Robot Wire Arc Additive Manufacturing (WAAM)","cited_arxiv_id":"2408.04677","evidence_quote":"Provides the multi-robot motion planning and software architecture that coordinates the three robots."},{"cited_title":"Hand-eye calibration for 2D laser profile scanners using straight edges of common objects,","cited_arxiv_id":null,"evidence_quote":"Hand-eye calibration for the 2D laser profile scanner, enabling point cloud registration."},{"cited_title":"A method for registration of 3-D shapes,","cited_arxiv_id":null,"evidence_quote":"ICP registration used to compare printed part to CAD model for accuracy evaluation."}],"review_version":1}