{"id":"241bd471-60aa-44a7-a59e-3469152a5c7f","arxiv_id":"2412.00396","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Distributed arm-mounted ToF depth sensors plus a transformer imitation policy reduce collisions and improve success in humanoid collision avoidance compared to external cameras and cuRobo.","lead":"This paper introduces ARMOR, a set of small time-of-flight depth sensors mounted on a humanoid robot's arms plus a transformer-based motion policy trained on human motion data. The authors report fewer collisions and faster planning than head-mounted or external cameras and than the cuRobo planner, with a qualitative real-robot demonstration.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table I(b)'s cuRobo comparison uses an ambiguous filter: success measured only where cuRobo solved would leave little room for a 16.9% improvement, while success measured on the full test set contradicts the stated filtering procedure.","rationale":"The reader's weakest assumption targets the obstacle-generation bias that favors arm-mounted sensors, which is a real generalization concern. However, I see a more immediate validity issue in the cuRobo comparison: the text and Table I(b) disagree about which test sequences contribute to the success metric. The paper's strongest quantitative claims are percentage improvements over baselines; if the baseline denominator is ambiguous, the numbers cannot be audited. This aligns with the reader's mention of 'evaluation filtering that could alter the baseline comparison,' so my agreement is partial. A conditional verdict remains appropriate: the concern can be settled by a clear re-analysis and reporting fix, and the hardware/perception contribution is not invalidated. I do not see evidence of bad faith; the issue is incomplete specification. The paper does have real strengths: a concrete hardware design, a substantial AMASS-based training pipeline, and a real-robot deployment, though the deployment is qualitative rather than measured.","tokens_in":9401,"tokens_out":7972,"duration_ms":86546,"concrete_test":"Recompute Table I(b) with explicit, auditable denominators. For the full 22,280 evaluation sequences and for the cuRobo-solvable subset separately, report: cuRobo collision count, cuRobo success count, ACT-Depth collision count, ACT-Depth success count, and the exact filtering rule used for each metric. If the 16.9% success improvement is not reproduced on the full set, or if cuRobo's success on the filtered subset is already at ceiling, the central IL-beats-sampling claim is unsupported. Release the evaluation split and filtering code so subset sizes and sequence IDs can be checked.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing problem is the denominator of the IL-vs-sampling comparison in Table I(b). Section IV.B states that cuRobo failed to find a solution on 64% of the evaluation data, and then 'we only use the sequences where cuRobo was able to find a solution when comparing the collision avoidance results... for fairness.' Yet Table I(b) reports ACT-Depth as having 16.9% higher success than cuRobo. Two readings are possible. First, if the success metric is restricted to the cuRobo-solvable subset, cuRobo's success rate should be close to 100% by construction, leaving no room for ACT-Depth to be 16.9% better. Second, if the success metric is computed over the full 22,280-sequence test set, the text's filtering description applies only to the collision comparison, and the two metrics use different denominators. Either way, the headline margin '31.6% fewer collisions, 16.9% higher success' is not well-defined from the paper as written. This is more direct than the benchmark-bias concern: it affects the validity of the reported baseline margin, not merely its generalization. The paper also provides no error bars or confidence intervals, but the undefined denominator is the first issue to resolve.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces ARMOR, a distributed egocentric perception system for humanoid robots based on small ToF lidar sensors mounted on the arms, together with ARMOR-Policy, a transformer-based imitation learning policy for collision avoidance trained on AMASS-derived human motion data. The central claim is that ARMOR perception is superior to head-mounted and externally mounted depth cameras, and that the learned policy is both safer and faster than the sampling-based planner cuRobo. The paper also reports a real-robot deployment on the GR1 humanoid.","tokens_in":9713,"tokens_out":7492,"duration_ms":67567,"significance":"If substantiated, the contribution is practically significant: a low-cost, low-power distributed sensor constellation that reduces occlusion for humanoid arm planning, combined with a data-driven policy, could be a compelling alternative to external cameras or computationally heavy sampling planners. The use of AMASS motion data to generate diverse expert demonstrations is a reasonable design choice, and the real-hardware deployment is a useful proof of concept. However, the evaluation protocol has several issues—the composition of the test set, an undefined comparison denominator for the cuRobo baseline, and the absence of error bars—that prevent the reported improvement margins from being taken at face value. With a corrected evaluation, the system would be a meaningful contribution to the field.","major_comments":[{"comment":"The text states that cuRobo failed to find a solution on 64% of the evaluation data, and that 'we only use the sequences where cuRobo was able to find a solution when comparing the collision avoidance results... for fairness.' This filtering is described only for the collision results, but Table I(b) reports a 16.9% higher success rate for ACT-Depth relative to cuRobo. If success is evaluated on the cuRobo-solved subset, cuRobo's success rate should be near 100% by construction, leaving little room for a 16.9% improvement. If success is evaluated on the full 22,280-sequence test set, then the collision and success metrics use different denominators, and the reported improvement is not well-defined. Please report absolute collision counts and success rates for both policies on the same test set, or clearly state the denominator for each metric.","section":"§IV.B, Table I(b)"},{"comment":"The 22,280 test sequences are described as a filtered subset of the 66,840 validation instances ('validated them on another 66,840 instances... This yields 22,280 motion sequences for testing'). If the same instances were used for validation (e.g., early stopping) and final evaluation, the reported success rates are optimistic and do not reflect generalization. The paper should either use a strictly disjoint test set or explicitly state that the validation set was never used for model selection. Additionally, the filtering criterion 'where a solution exists' needs a precise definition: a solution for which planner/policy, and what happens to the excluded sequences in each metric?","section":"§IV.A"},{"comment":"In the data-generation scheme, obstacles are placed tightly around the re-targeted AMASS trajectory (Section III.D, Figure 5). Since the test sequences come from the same generator, obstacles are by construction close to the arm-mounted sensors, while head-height exocentric cameras may be occluded by the robot's body or by clutter. This protocol may systematically favor ARMOR over exocentric perception, independent of the sensor hardware. A more informative evaluation would include held-out environments with obstacles not derived from the expert trajectory, and would report whether the ARMOR advantage persists when obstacles are placed away from the arms or when occlusions are reduced.","section":"§III.D, §IV.A"},{"comment":"The paper states that in emergency-stop demonstrations 'the last goal position [is] a random location lying inside of an obstacle, which always leads the arm to collide,' yet these trajectories are included in the demonstration data used to train an imitation learning policy for collision avoidance. The manuscript does not explain how these colliding demonstrations are used: are they treated as positive examples, negative examples, or filtered out? If they are used as positive demonstrations, the policy would be trained to imitate colliding motions; if they are filtered out, the description should say so. Please clarify the role of emergency-stop data in the training set.","section":"§III.D"}],"minor_comments":[{"comment":"The abstract and Section IV.C state that the source code, hardware description, and CAD files will be released, but no link is provided in the submitted manuscript. Please include the actual repository/URL or a clear statement of availability in the final version.","section":"Abstract, §IV.C"},{"comment":"The caption contains a typo: 'VL35L5CX' should be 'VL53L5CX'.","section":"Figure 6 caption"},{"comment":"The paper uses '86 hours' and '86.6 hours' interchangeably for the training data; please standardize the number.","section":"§I, §IV.A"},{"comment":"The table reports only percentage improvements without absolute values; adding absolute counts (e.g., number of collisions, number of successes out of N) would make the results interpretable.","section":"Table I"},{"comment":"The description of point-cloud pruning for cuRobo is ambiguous when using ARMOR perception: 40 sensors at 8x8 resolution produce only 2,560 points, so the 10,000-point cap and the 3.7M points likely refer to the exocentric camera setup. Please clarify which perception configuration the pruning applies to.","section":"§IV.B"},{"comment":"The paper says 'we also simulate the sensor noise' but does not give the noise model used for the VL53L5CX or the D435; specify the noise parameters for reproducibility.","section":"§IV.A"}],"recommendation":"major_revision","confidential_remarks":"The paper is a viable systems contribution, but the evaluation protocol is the main obstacle. The undefined denominator in the cuRobo comparison and the use of the validation set as the test set are concrete, fixable issues. I would also ask the editor to require the code and hardware links before publication, since the manuscript currently withholds them."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper has a genuinely interesting idea—arm-mounted, low-res ToF arrays for humanoid collision avoidance—but the main numerical claim against cuRobo is not well-defined from the text. The system concept is worth engaging; the evaluation needs rework.\n\nThe good stuff first. The authors place 40 cheap 8x8 VL53L5CX ToF sensors across the GR1 arms, pair them with an ACT-style transformer policy, and generate training data by retargeting 86 hours of AMASS human motion into tight obstacle scenarios. The intuition is solid: occlusion matters most near the end effector, and distributed low-fidelity depth might beat a few high-fidelity head/external cameras. Real-hardware deployment with 28 sensors gives the work credibility. This is a reasonable systems contribution, and the novelty is real even if incremental.\n\nThe soft spot is load-bearing. Section IV.B says cuRobo failed on 64% of evaluation sequences and that 'we only use the sequences where cuRobo was able to find a solution when comparing the collision avoidance results.' Then Table I(b) reports a 16.9% success improvement for ACT-Depth over cuRobo. If success is computed on the cuRobo-solvable subset, cuRobo's success rate is by construction near ceiling, so 16.9% relative improvement is implausible. If success is computed on the full test set, then collision and success use different denominators, and the comparison is not apples-to-apples. The paper does not say which. This is exactly the kind of thing a referee would catch, and it makes the headline '31.6% fewer collisions, 16.9% higher success' unverifiable as written.\n\nSecondary issues: no error bars anywhere, and the evaluation protocol generates obstacles tightly around the demonstrated trajectories, which may systematically favor arm-mounted sensors over exocentric cameras. Section III.D and IV.A confirm the generation strategy. Code, CAD, and video are promised but absent from this version. These are fixable.\n\nWho this is for: anyone working on egocentric perception for humanoid manipulation or learned collision avoidance. The ARMOR concept is worth having on your radar, but the quantitative results should be treated with caution until the evaluation is clarified. This deserves a serious referee—the core idea is sound and the paper is not inherently flawed—but it needs a major revision on the experimental reporting. I'd advise sending it out, with instructions to the authors to specify denominators, add error bars, and release artifacts.","headline":"ARMOR has a genuinely useful systems idea—distributed arm-mounted ToF sensors for humanoid collision avoidance—but the headline cuRobo comparison has an undefined denominator and needs a careful rewrite.","tokens_in":10198,"tokens_out":3761,"would_cite":true,"duration_ms":36381,"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":"ARMOR claims that distributed arm-mounted time-of-flight sensors, paired with a transformer imitation policy, cut humanoid collisions by 63.7% over head/external cameras and run 26x faster than sampling-based planning.","keywords":["egocentric perception","time-of-flight sensors","humanoid robot","collision avoidance","imitation learning","transformer policy","motion planning","depth sensor constellation"],"falsifier":"Run the same ARMOR and four-camera perceptual setups on a test set where obstacles are placed freely at various heights rather than generated around the demonstration trajectories, and compare collision counts and success rates; if ARMOR's advantage shrinks or reverses, the tight-trajectory data generation was the source of the gain.","tokens_in":9215,"feed_emoji":"🤖","tokens_out":8206,"duration_ms":68667,"temperature":0.7,"pith_summary":"This paper argues that the standard way of equipping a humanoid robot with vision — a dense camera on the head or mounted outside the body — is the wrong trade for collision avoidance. It introduces ARMOR, a constellation of forty small, low-cost time-of-flight depth sensors distributed across the robot's arms, each returning a coarse 8-by-8 depth image in its own frame, and pairs this perception layer with a transformer policy trained by imitation on human motion data. In simulation, feeding ARMOR's arm-mounted point clouds to the same policy cuts collisions by 63.7% and raises task success by 78.7% relative to feeding it four dense head-height cameras. The paper also reports that the learned policy beats a sampling-based motion planner with 31.6% fewer collisions, 16.9% higher success, and about 26 times less computation. The claim, if it holds, is that sparse, distributed, low-resolution perception placed where collisions actually happen can outperform high-resolution perception placed where the robot can see its hands.","feed_headline":"Arm-mounted depth sensors cut robot collisions by 64%","feed_subtitle":"A transformer policy trained on human motions avoids obstacles 26x faster than a sampling planner.","key_machinery":"The carrying object is ARMOR's distributed sensor constellation: forty time-of-flight zone-array lidars mounted twenty per arm, each producing an 8x8 depth image with a 63-degree field of view and streaming at 15 Hz. The software side is ARMOR-Policy, a transformer encoder-decoder built on an action-chunking architecture. It ingests the forty ego-frame depth images through monochannel ResNet encoders, together with current and goal joint positions and a latent style variable z that encodes trajectory diversity, and outputs action sequences for the 14 arm degrees of freedom. At inference the policy samples multiple candidate trajectories in parallel by varying z and selects the one that minimizes the robot-to-point-cloud distance evaluated with a signed distance function.","core_discovery":"The central claim is that egocentric, distributed, low-resolution depth sensing on a humanoid's arms is a better perceptual substrate for collision avoidance than centralized high-resolution cameras, and that a transformer trained to imitate human motions can exploit it. In the paper's dense-obstacle simulation benchmark built from retargeted human arm motions, ARMOR-Policy reduces collisions by 63.7% and improves success rate by 78.7% over the same policy running on a four-camera exocentric setup that includes a head-mounted camera. Compared with a sampling-based motion planning expert using the same ARMOR perception, the neural policy shows 31.6% fewer collisions, 16.9% higher success, and about 26 times lower computation time. The system was also deployed on a real humanoid with 28 ToF lidars, updating its trajectory at 15 Hz.","pith_inferences":["A natural boundary of the claim is open environments: when obstacles are visible from a head-mounted camera and far from the arms, the exocentric setup may catch up, so the reported advantage should be expected to shrink outside cluttered, arm-near scenes.","An ablation that varies the number and mounting locations of the ToF sensors would reveal where the gains come from and how few sensors are still sufficient.","The same distributed egocentric sensing could apply to other manipulator morphologies, such as mobile manipulators or legged robots with arms, where head-mounted cameras share the same occlusion blind spots.","Because the input images are only 8x8, the policy could likely be distilled into a much smaller network that runs on embedded hardware, enabling closed-loop avoidance onboard without a workstation GPU."],"forward_implications":["Humanoid robots could rely on body-mounted sensors instead of external cameras for collision avoidance, supporting mobile, untethered operation in dense spaces.","The roughly 26x reduction in planning latency enables reactive avoidance at the sensor's 15 Hz update rate, which is fast enough for dynamic environments.","Because the policy consumes generic ego-frame depth images, the same trained policy may transfer to different arm geometries or sensor layouts without retraining.","The low cost and low profile of the sensors make whole-body coverage economically feasible, effectively giving the robot a coarse artificial skin.","Sampling many latent trajectories at inference and picking the one with largest clearance adds a safety margin that would be expensive to obtain with a single deterministic plan."],"supporting_citations":[{"why":"Supplies the large motion-capture corpus of human arm motions used as expert demonstrations for imitation learning.","marker":"[20]"},{"why":"Provides the action-chunking transformer architecture that ARMOR-Policy's encoder-decoder is based on.","marker":"[34]"},{"why":"Defines the sampling-based motion planning expert used as the baseline for collision, success, and latency comparisons.","marker":"[29]"},{"why":"Introduces exocentric camera baselines for neural motion planning and the four-D435 setup ARMOR is compared against.","marker":"[3]"},{"why":"Documents the SparkFun VL53L5CX time-of-flight zone-array sensor that the distributed constellation is built from.","marker":"[26]"},{"why":"Establishes the imitation-learning approach for neural motion planning that ARMOR-Policy extends with human-motion data.","marker":"[5]"}],"fun_headline_variants":["Arm sensors slash humanoid crashes by 64%","Humanoid avoids obstacles with arm cameras, plans 26x faster","Egocentric arm depth sensing beats head cameras for collision avoidance","Transformer policy on arm sensors cuts robot collision rate 64%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark generates obstacles tightly around the recorded human arm motions, so the arm-mounted sensors almost always see the obstacles directly while head-height cameras are often occluded.","fun_headline_variants_meta":{"raw":{"variants":["Arm sensors slash humanoid crashes by 64%","Humanoid avoids obstacles with arm cameras, plans 26x faster","Egocentric arm depth sensing beats head cameras for collision avoidance","Transformer policy on arm sensors cuts robot collision rate 64%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000374,"raw_usage":{"total_tokens":1991,"prompt_tokens":937,"completion_tokens":1054,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":985}},"tokens_in":553,"tokens_out":1054,"duration_ms":10027,"temperature":1.0,"reasoning_tokens":985,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:25:29.991652+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same ARMOR and four-camera perceptual setups on a test set where obstacles are placed freely at various heights rather than generated around the demonstration trajectories, and compare collision counts and success rates; if ARMOR's advantage shrinks or reverses, the tight-trajectory data generation was the source of the gain.","supporting_citations":[{"cited_title":"Troje, Gerard Pons- Moll, and Michael J","cited_arxiv_id":null,"evidence_quote":"Supplies the large motion-capture corpus of human arm motions used as expert demonstrations for imitation learning."},{"cited_title":"Curobo: Parallelized collision-free robot motion generation","cited_arxiv_id":null,"evidence_quote":"Defines the sampling-based motion planning expert used as the baseline for collision, success, and latency comparisons."},{"cited_title":"Neural mp: A generalist neural motion planner","cited_arxiv_id":null,"evidence_quote":"Introduces exocentric camera baselines for neural motion planning and the four-D435 setup ARMOR is compared against."},{"cited_title":"Qwiic mini tof imager - vl53l5cx, 2024","cited_arxiv_id":null,"evidence_quote":"Documents the SparkFun VL53L5CX time-of-flight zone-array sensor that the distributed constellation is built from."},{"cited_title":"Motion policy networks","cited_arxiv_id":null,"evidence_quote":"Establishes the imitation-learning approach for neural motion planning that ARMOR-Policy extends with human-motion data."}],"review_version":1}