{"id":"849f20e0-d64b-4e74-99f6-dfe4f57f31ca","arxiv_id":"2505.11848","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A transformer trained on proprioceptive histories estimates the positions, sizes, and mobility of up to three rectangular obstacles, including occluded ones, from a quadruped's contact interactions alone.","lead":"PROBE is a transformer that maps a quadruped robot's joint torques, motions, and pose into live estimates of nearby rectangular obstacles, including boxes hidden behind other boxes, using no cameras. It shows how much a robot can infer about its surroundings from touch and motion alone, which matters when vision is blocked by smoke, darkness, or rubble.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed SE(2) obstacle pose estimation is only evaluated under fixed zero initial orientation; static obstacle orientation error is explicitly omitted, so the central claim of predicting arbitrary planar box poses is unsupported.","rationale":"The most load-bearing concern is the orientation restriction, because it directly undercuts the abstract and introduction's explicit promise of pose estimation in SE(2). The experiments never require the model to estimate a static obstacle's orientation, and the model is never tested on initially rotated obstacles. This is a real gap: contact geometry, push directions, and the proprioceptive signatures of collisions all change with orientation, so a network trained only on axis-aligned boxes may fail on rotated ones. I considered whether the external-pose input (ArUco/motion capture in real trials) is a stronger concern, but the paper frames body pose as part of the proprioceptive input and notes it can be replaced by onboard localization, so it is less directly tied to the claimed contribution than the SE(2) generalization gap. The absence of baselines and error bars is a weakness but does not by itself invalidate the absolute quantitative results. The conclusion's stated limitation about box-shaped objects does not mention the orientation restriction, so the concern is not pre-acknowledged in the paper. The reader's weakest assumption captures this same issue, and the recommended verdict remains CONDITIONAL: the central claim should be accepted conditional on demonstrating performance over a nontrivial range of obstacle orientations.","tokens_in":10258,"tokens_out":6471,"duration_ms":76214,"concrete_test":"Re-run the Table I simulation evaluation with obstacle initial orientations sampled uniformly in [-pi/4, pi/4] (or, more strongly, uniformly over [-pi, pi]) using the same trained ORM, and report IoU and theta error separately for static and movable obstacles by orientation bin. If static IoU or theta error degrades substantially for nonzero orientations, the model has not learned general SE(2) pose estimation and the claim should be narrowed to axis-aligned boxes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V (Environment Setup) states that during training, obstacles are 'placed in random SE(2) configurations with their orientations fixed to 0 with respect to the robot's initial orientation.' The output representation in Section III explicitly includes theta_i in SE(2), and the abstract and introduction claim that PROBE predicts obstacle 'poses in SE(2).' Yet every reported simulation and real-robot benchmark inherits this zero-orientation initial condition. Table I goes further and omits orientation error for static obstacles, saying 'Since the orientation of the static obstacles is fixed across the different benchmarks, their orientation error is omitted.' This is an explicit acknowledgment that static obstacle orientation is never varied, never predicted in a nontrivial sense, and never evaluated. Movable obstacles may acquire nonzero orientation through being pushed, so some theta variation exists dynamically, but the initial and static cases are still axis-aligned. The central claim therefore overstates what the evidence supports: the ORM may have learned an axis-aligned prior rather than a general SE(2) pose estimator, and the claimed capability to detect unseen rotated rectangular obstacles is not demonstrated. This is a validation and generalization gap, not an internal inconsistency, and it is the same weakest assumption the reader identified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PROBE, a Transformer-based obstacle reconstruction module that takes a history of proprioceptive signals (joint positions, velocities, torques, and robot pose) from a Unitree Go1 quadruped and predicts the presence, dimensions, pose, and static/movable status of rectangular obstacles in a 2D workspace, including obstacles occluded by movable obstacles. The method is trained in Isaac Gym on trajectories generated by learned navigation policies and evaluated in simulation (1000 trials per difficulty) and on a real robot (20 Easy, 20 Medium, 5 Hard trials). The paper reports IoU and absolute error metrics, and an ablation on input modalities.","tokens_in":10533,"tokens_out":4595,"duration_ms":42328,"significance":"If the claims hold, this is a novel contribution: using only proprioception to reconstruct a 2D obstacle map, including nested occluded obstacles, on a legged robot without vision. The paper provides a clear problem formulation, a scalable simulation pipeline, and real-robot validation. The architecture is straightforward, and the ablation gives evidence about which inputs matter. However, the evaluation has gaps—most notably the orientation issue and the lack of baselines—that limit the strength of the central claims.","major_comments":[{"comment":"The abstract and introduction claim that PROBE predicts obstacle poses in SE(2), but Section V states that during training obstacles are placed in random SE(2) configurations with their orientations fixed to 0 with respect to the robot's initial orientation. Table I then omits the orientation error for static obstacles with the note that their orientation is fixed across benchmarks. This is an explicit acknowledgment that static obstacle orientation is never varied or evaluated. The reported theta errors for movable obstacles (0.198 to 0.214 in Table I) come only from dynamic reorientation during pushing, which does not exercise arbitrary initial orientations. The manuscript should either add experiments with rotated obstacles (both static and movable) and report static orientation error, or restrict the claims to axis-aligned obstacles. As written, the general SE(2) pose-estimation claim is not supported by the evidence.","section":"V, Table I"},{"comment":"The real-robot evaluation of the Hard benchmark rests on only 5 trials (Table II), and no variance, confidence intervals, or per-trial results are reported for any of the simulation or real-robot results. The claim of real-world effectiveness under the Hard scenario is therefore fragile. Please report standard deviations or ranges, and ideally increase the number of Hard trials.","section":"VI.B, Table II"},{"comment":"No baseline or comparison method is presented. While the ablation in Fig. 6 studies input modalities and is useful, it is conducted only for Nmax=2 and the evaluation set size is not specified. Without an external baseline (e.g., a contact-window heuristic or a non-transformer sequence model), it is hard to judge whether the learned Transformer mapping contributes beyond the information available at the final contact time step. The authors should either provide such a baseline or explicitly justify why no existing method is directly comparable.","section":"VI"}],"minor_comments":[{"comment":"The phrase 'provides an novel method' should read 'provides a novel method'.","section":"IV.E"},{"comment":"The notation for the output set, \\hat{O}_t = \\{\\hat{O}^t_i\\}_{i=N}^{i=1}, uses N for the number of obstacles while the surrounding text uses n. Please unify the notation.","section":"IV.D"},{"comment":"The environment dimensions wenv and lenv appear in Fig. 2 but are not defined in the text; please define them explicitly.","section":"V"},{"comment":"The sentence 'The reported reconstruction for the movable obstacle is surprisingly more accurate in the Medium and Hard scenarios' is speculative; please provide a more precise explanation or remove the word 'surprisingly'.","section":"VI.A"},{"comment":"The dataset curation section mentions pruning trajectories based on contact mode frequency, but the pruning rule and the resulting dataset composition are not described; please add details.","section":"IV.C"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is interesting and the real-robot demonstration is valuable, but the SE(2) orientation gap is a substantive validation issue that should be addressed before publication. The lack of baselines may also be a concern for the target journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: PROBE is the first thing I've seen that runs a transformer over a quadruped's proprioceptive history and outputs box obstacle parameters, including the case where a movable box hides a static one behind it. That nested-occlusion capability is the real novelty, and they back it with both Isaac Gym and a physical Go1. Credit where due: the system integrates a trained locomotion policy, a navigation policy, and a supervised reconstruction module; the data curation discussion (policy bias, contact mode bias) shows the authors understand the failure modes; and the conclusion honestly lists box-only shape and lack of other physical properties as limitations.\n\nThe soft spots. The biggest one is the one the authors half-acknowledge: Section V says all obstacles are placed with orientation fixed to zero relative to the robot's initial frame, and Table I omits orientation error for static obstacles \"since the orientation of the static obstacles is fixed across the different benchmarks.\" So the abstract and intro claim of predicting poses in SE(2) is not supported for rotated rectangles. The network may have learned an axis-aligned prior. This is a validation gap, not an internal contradiction, but it needs to be fixed or the claims need to be narrowed to axis-aligned boxes. Everything else is smaller: no baseline comparisons against prior force-based mapping (e.g., inverse physics) or even a simple odometry-tactile heuristic; no error bars anywhere; only 5 real Hard trials; the real occluded-static IoU is around 0.2, which is weak; the ablation is only on N_max=2 and the test set isn't specified. None of these are fatal, but they collectively mean the empirical case is thinner than the prose suggests.\n\nThe citation pattern looks fine; related work is on-point. I don't see circularity — the IoU numbers are measured on held-out environments, and using your own navigation policy to generate training data is standard.\n\nBottom line: this paper deserves serious review, not a desk reject, but a reviewer should push for either rotated-obstacle experiments or a rewritten claim, plus baselines and error bars. I'd bring it to our reading group as an example of a good idea with a too-broad claim.","headline":"Useful new system for contact-based obstacle mapping on a legged robot, but the SE(2) pose claim is only tested for axis-aligned boxes and the evaluation needs more rigor.","tokens_in":11039,"tokens_out":1793,"would_cite":true,"duration_ms":18157,"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 quadruped can reconstruct a 2D map of box obstacles, including fully occluded ones, from joint torques, joint states, and body pose alone.","keywords":["proprioceptive perception","obstacle reconstruction","legged robot navigation","transformer","contact sensing","movable obstacles","occlusion","scene reconstruction"],"falsifier":"Place a single movable box at a 45-degree angle to the robot's starting heading in the real corridor and let the robot navigate and push it: the ORM, trained only on zero-rotation boxes, either tracks the true angle with IoU near the reported movable-obstacle range or the reconstruction collapses, which settles whether the SE(2) claim generalizes or was carried by axis-aligned training. A second check replays one recorded real trajectory through the ORM and counts how often the hidden static box behind the pushed movable one is recovered, since that occlusion capability is the paper's strongest claim.","tokens_in":10126,"feed_emoji":"🐕","tokens_out":15264,"duration_ms":135111,"temperature":0.7,"pith_summary":"PROBE claims that a quadruped robot can reconstruct a partial 2D map of box-shaped obstacles — each box's center, dimensions, rotation, and whether it is fixed or can be pushed — using nothing but its own body as the sensor: a history of joint positions, joint velocities, applied torques, and body pose, with no camera or other exteroceptive input. The strongest part of the claim is occlusion handling: when the robot pushes a movable box into a static one, the network reports the hidden static box and its dimensions, something vision alone cannot do. To back this up, the paper trains a causal Transformer on 180,000 simulated exploration trajectories and evaluates it on a real quadruped, with final-contact rotated intersection-over-union values between roughly 0.33 and 0.50 in simulation and between 0.20 and 0.45 on the physical robot. If the claim stands, legged robots gain a mapping modality that works exactly where cameras fail: darkness, smoke, and scenes where the object of interest sits behind another object.","feed_headline":"Maps hidden boxes using only touch and torque","feed_subtitle":"A transformer turns joint and torque history into box positions, sizes, and hidden obstacles behind movable ones.","key_machinery":"The central object is the Obstacle Reconstruction Module (ORM): a causal Transformer encoder followed by a two-layer MLP decoder that reads a history of proprioception vectors — joint positions, joint velocities, commanded joint torques, and the robot's SE(2) pose — and outputs one parameterized box per obstacle, (I_static, x, y, θ, w, l). Two design decisions carry the argument. First, the contact window: ground-truth labels for each obstacle are masked except over the interval during which the robot or a box pushed by the robot is physically touching it, and the training loss is a weighted sum of binary cross-entropy for the contact and static flags plus mean squared error for pose and dimensions, so the network is supervised exactly on the information-rich phase of each interaction. Second, the data-generation stack: a PPO-trained high-level navigation policy commands a student-teacher locomotion policy, and the resulting 180,000 curated trajectories span contact modes from no contact to direct contact with a movable box that indirectly pushes a hidden static box, which is what lets the ORM learn nested occlusion.","core_discovery":"The paper's central discovery is that a supervised Transformer can map a quadruped's proprioception history into an actionable scene description with no visual input. The Obstacle Reconstruction Module (ORM) outputs, for each obstacle, a tuple giving a static-or-movable flag, an SE(2) pose (a 2D position plus a rotation angle), and the box's width and length, and it is trained to make each prediction only during that obstacle's contact window — the interval in which the robot or a pushed neighboring box is mechanically pressing on it. At the end of the contact window the final estimate is scored with rotated intersection-over-union between the predicted and ground-truth box geometry, and the reported means land at 0.33-0.50 for simulated scenarios of one to three obstacles and at 0.20-0.45 on the real quadruped. The paper also demonstrates the nested-interaction case in which a static box fully hidden behind a movable one is reconstructed only after the movable box has been pushed against it, which is the capability that most distinguishes proprioceptive reconstruction from any camera-based map.","pith_inferences":["All training environments fix obstacle orientation to zero, so the SE(2) claim is only demonstrated for axis-aligned boxes; the immediate test is to rotate a box 15-45 degrees and check whether the network's rotation estimate holds, and if it collapses, orientation augmentation of the training set becomes necessary.","The contact-window design suggests reconstruction quality tracks the amount of sustained contact the navigation policy produces, so an exploration policy that slides along box faces rather than bumping corners should raise final IoU more than any network change.","The rectangular parameterization caps what the network can express; swapping the decoder for an occupancy grid or an implicit shape representation would carry the same proprioception pipeline to arbitrary obstacle shapes, at the cost of the compact interpretable box output.","The proprioceptive map is best viewed as a complement to vision rather than a replacement: a fused system could trust the ORM exactly when the camera is occluded or the scene is dark, and cross-modal disagreement could itself flag which obstacles are movable."],"forward_implications":["A quadruped can keep a working 2D obstacle map in darkness, smoke, or rubble where cameras are useless, using only its own joint torques, joint states, and body pose.","Because a pushed movable box reveals obstacles hidden behind it, the robot can plan around what it has not yet seen directly, which is the information required for navigation among movable obstacles.","The ORM produces estimates online during navigation and refines them as contact accumulates, so a planner can act on each box as soon as its contact window ends rather than waiting for a post-run map.","Sim-to-real transfer is achieved without retraining: real-robot reconstructions follow the same accuracy trends as simulation, with the degradation attributed by the paper to mass, friction, and shape differences between simulated and physical boxes."],"supporting_citations":[{"why":"Supplies the student-teacher adaptation framework used to train the low-level locomotion controller so it transfers across terrain and contact conditions.","marker":"[24]"},{"why":"Provides the neural-network locomotion policy that the navigation policy commands and that physically generates the contacts the reconstruction module reads.","marker":"[27]"},{"why":"Presents the proximal policy optimization algorithm used to train both the locomotion and navigation policies.","marker":"[28]"},{"why":"Defines the Transformer encoder architecture that the obstacle reconstruction module is built on.","marker":"[29]"},{"why":"Defines the navigation-among-movable-obstacles problem that the reconstruction output is meant to serve.","marker":"[8]"},{"why":"Frames the blindfolded-traveler variant of navigation with contact feedback only, against which this paper's fixed-plus-movable obstacle setup is contrasted.","marker":"[20]"},{"why":"Represents the prior physics-based inverse-reasoning approach that searches large shape spaces in clutter, which PROBE's learned direct mapping aims to replace.","marker":"[10]"}],"fun_headline_variants":["Robot maps hidden boxes using only its own motion","Blind robot detects and sizes boxes from touch alone","Robot pushes obstacles to find hidden ones without vision","From torque to map: robot estimates hidden boxes by feel","Proprioception alone reveals hidden boxes in cluttered scenes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result currently rests on the assumption that every obstacle is an axis-aligned rectangle: the training environments fix all obstacle orientations to zero and no rotated box is ever evaluated, so the claimed full pose estimation is unverified for any other angle.","fun_headline_variants_meta":{"raw":{"variants":["Robot maps hidden boxes using only its own motion","Blind robot detects and sizes boxes from touch alone","Robot pushes obstacles to find hidden ones without vision","From torque to map: robot estimates hidden boxes by feel","Proprioception alone reveals hidden boxes in cluttered scenes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000651,"raw_usage":{"total_tokens":2967,"prompt_tokens":905,"completion_tokens":2062,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":1985}},"tokens_in":521,"tokens_out":2062,"duration_ms":17215,"temperature":1.0,"reasoning_tokens":1985,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:46:10.382353+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Place a single movable box at a 45-degree angle to the robot's starting heading in the real corridor and let the robot navigate and push it: the ORM, trained only on zero-rotation boxes, either tracks the true angle with IoU near the reported movable-obstacle range or the reconstruction collapses, which settles whether the SE(2) claim generalizes or was carried by axis-aligned training. A second check replays one recorded real trajectory through the ORM and counts how often the hidden static box behind the pushed movable one is recovered, since that occlusion capability is the paper's strongest claim.","supporting_citations":[{"cited_title":"Walk these ways: Tuning robot control for generalization with multiplicity of behavior,","cited_arxiv_id":null,"evidence_quote":"Provides the neural-network locomotion policy that the navigation policy commands and that physically generates the contacts the reconstruction module reads."},{"cited_title":"Navigation among movable obstacles,","cited_arxiv_id":null,"evidence_quote":"Defines the navigation-among-movable-obstacles problem that the reconstruction output is meant to serve."},{"cited_title":"The blindfolded robot: A bayesian approach to planning with contact feedback,","cited_arxiv_id":null,"evidence_quote":"Frames the blindfolded-traveler variant of navigation with contact feedback only, against which this paper's fixed-plus-movable obstacle setup is contrasted."},{"cited_title":"Inferring 3d shapes of unknown rigid objects in clutter through inverse physics reasoning,","cited_arxiv_id":null,"evidence_quote":"Represents the prior physics-based inverse-reasoning approach that searches large shape spaces in clutter, which PROBE's learned direct mapping aims to replace."}],"review_version":1}