{"id":"da234200-d367-4025-a7c9-f7eeb6d9e903","arxiv_id":"2412.13621","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":12,"one_line_summary":"An RL-trained quadruped policy, using bidirectional scandots as privileged information and a centerline reward, crosses narrow pipes with obstacles in simulation and on a Unitree Go2.","lead":"A reinforcement learning pipeline lets a quadruped robot crawl through narrow pipes, using simulated bidirectional height scans as privileged training information and a depth camera at deployment. The authors demonstrate the policy on a Unitree Go2 crossing PVC pipes with radius as small as 0.192 meters, though with lower real-world success than in simulation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The only real-world obstacle result in Table V (Pipe C, 0.192 m, with obstacles) reports a 46.5% traversing percentage with no trial count, so the central claim of crossing pipes with unexpected obstacles is not actually demonstrated.","rationale":"The reader's conditional verdict is based primarily on the sim-to-real gap and the lack of trial counts in the real experiments. My analysis agrees with that direction but identifies a more specific and more damaging inconsistency: the only real-world obstacle row in Table V reports 46.5% traversing, which contradicts the claim of successful obstacle crossing and is unexplained. This is not a question of simulation fidelity or baseline comparisons; it is an internal mismatch between the headline claim and the paper's own table. The concern is load-bearing because the abstract explicitly highlights obstacle handling as a demonstrated capability. The concrete test is deliberately narrow: clarify the metric and provide a complete trial log with denominators. If the obstacle trials fail to show a meaningful success rate, the central claim should be weakened to obstacle-free pipe crossing only; if they show good success, the conditional verdict can stand. I keep the reader's UNCHANGED verdict because the paper already deserves CONDITIONAL status on other validation grounds, and my concern strengthens that condition rather than overturning the plausible feasibility result.","tokens_in":9494,"tokens_out":2862,"duration_ms":27711,"concrete_test":"Run a logged real-world obstacle evaluation on the same PVC pipe setup as Table V. Define 'Traversing Percentage' explicitly, then run at least 20 independent trials on Pipe C (radius 0.192 m, length 2 m, with random obstacles) and report: total trials, number of successful full traversals, distances reached on failures, completion times, and the exact binomial 95% confidence interval for the success rate. Also run at least 10 trials without obstacles as a control. If the metric is success rate, require the lower confidence bound to exceed an a-priori threshold such as 50%; if the metric is fraction of pipe length, require each reported success to reach 100% traversal. This single check determines whether the 'unexpected obstacles' clause of the central claim is supported by real-world data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and introduction claim that the proposed RL method enables a quadruped to cross narrow pipes 'even with unexpected obstacles inside.' The only direct real-world evidence for this is Table V's obstacle row: Pipe C, radius 0.192 m, length 2 m, with random obstacles, 'Traversing Percentage' 46.5%. This is internally inconsistent with the surrounding text, which calls Table V 'the statistical results of the selected success real-world experiments' and concludes the robot 'can also deal with those circumstances which has obstacles inside.' If 46.5% is a success rate, then the obstacle task failed more than half the time; if it is the average fraction of the pipe traversed, then the robot did not complete the pipe with obstacles at all. Either way, there is no total trial count, no definition of the metric, no failure breakdown, and no raw logs. The simulation results are strong, but the real obstacle experiment is the bridge from simulation to the 'even with obstacles' part of the headline claim, and that bridge is a single, poorly specified number. The paper's own conclusion further concedes that the method 'may deteriorate when the visual input has a large noise or the robot is stuck by some unseen obstacles.' This concern does not invalidate the feasibility of pipe crossing without obstacles, but it does mean the obstacle-handling claim is not currently established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a three-stage reinforcement-learning curriculum that trains a Unitree Go2 policy to traverse narrow pipes. The method introduces a custom height-field pipe terrain (Eq. 1), a privileged bidirectional scandots observation (Section II-B), and a reward function combining velocity tracking, energy penalties, collision avoidance, and a centerline-distance penalty (Section II-C). Training is performed in Isaac Gym with PPO, and simulation results in Table IV show that the final-stage policy reaches 93.75% success on 0.2 m radius pipes with random obstacles. Real-world experiments on 2-3 m PVC pipes are reported in Table V, and the paper claims the robot can cross pipes even with obstacles inside.","tokens_in":9795,"tokens_out":5331,"duration_ms":50054,"significance":"If the claims are fully validated, this would be a useful application of learned legged locomotion to narrow-pipe inspection, a domain where classical in-pipe robots are highly specialized. The simulation results are internally consistent and show a clear curriculum effect: success on the hardest simulated condition rises from 0% (Stage 1) and 0% (Stage 2) to 93.75% (Stage 3) in Table IV, which is credible evidence of learning. The bidirectional scandots idea is plausible and the reward formulation is standard enough to be reproducible. However, the real-world evidence is currently too weak to support the headline claim, and the absence of any baselines or ablations means the paper's stated contributions are not isolated. I found no equation-level circularity in the reward design; the main risks are empirical and reporting-related.","major_comments":[{"comment":"The only real-world evidence for the headline claim of traversing pipes 'even with unexpected obstacles inside' is the Table V row for Pipe C with obstacles, which reports 'Traversing Percentage' 46.5% with no definition, no total trial count, and no failure breakdown. The reported average speed (0.078 m/s over 12 s) is numerically consistent with traversing 46.5% of the 2 m pipe (0.93 m), which suggests the metric is the average fraction of pipe length traversed rather than a success rate; under that reading the robot did not complete the obstacle pipe. The surrounding text also calls Table V the 'statistical results of the selected success real-world experiments,' implying data selection, which further undermines any success-rate interpretation. Since Section IV itself concedes that the method 'may deteriorate when the visual input has a large noise or the robot is stuck by some unseen obstacles,' the obstacle-handling claim needs additional real-world trials with complete denominators, a precise metric definition, and a breakdown of failure modes.","section":"Section III-D, Table V"},{"comment":"The real-world validation does not test the task as trained. The policy is trained and evaluated in simulation on 18 m pipes (Table IV), but the real pipes are only 2-3 m long, so the learned behavior is not shown to persist over the full training distance. The text states that the real success rate is 'much lower' than in simulation but gives no quantitative real success rate, and the physical factors (slippery PVC, unfixed pipe sway, noisy sensors) are only described qualitatively. Without a quantitative success rate or a complete trial log, the sim-to-real transfer claim is not established, especially because the real obstacle configuration is not specified in a way that can be matched to the simulated obstacle sampling.","section":"Section III-D"},{"comment":"No ablation or baseline is provided for the claimed contributions. Table IV compares the same method at different curriculum stages, but there is no comparison without the bidirectional scandots, without the centerline-distance penalty, with different reward weights, or against an existing quadruped locomotion method. Because the reward function in Section II-C combines several standard LeggedGym terms with new terms, the current experiments do not isolate which component enables pipe crossing, so the stated contributions are not yet validated.","section":"Section II-C and Table IV"}],"minor_comments":[{"comment":"The observation vector is declared o_t ∈ R^49, but the listed components sum to 3+2+3+3+12+12+12=47, and adding the privileged vector ξ_t ∈ R^296 gives R^343; please correct the dimension or the component list.","section":"Eq. (2) and Table I"},{"comment":"There are several typographical errors: 'deply' should be 'deploy' in the contribution list, 'trails' should be 'trials' in Section III-D, 'robot'current' should be 'robot's current' in Section II-B, and the capitalization in 'In the first stage, We introduce' is inconsistent.","section":"Section I and Section III-D"},{"comment":"The height-field formula is said to be 'unrelated to x,' which is fine, but the notation would be clearer if the domain of y and the meaning of z0 (height of the pipe centerline) were stated explicitly before the equation.","section":"Section II-A, Eq. (1)"},{"comment":"The column 'Traversing Percentage' is not defined anywhere in the text, and 'Completion Time' is ambiguous for the obstacle row; please define these metrics and state how they are computed over the trials.","section":"Table V"}],"recommendation":"major_revision","confidential_remarks":"The 'selected success' phrasing in Section III-D is the key weakness. I would ask the authors for the full trial-by-trial logs and the precise definition of 'Traversing Percentage' before any acceptance decision. The manuscript is otherwise within scope for a robotics venue, and the simulation results provide a credible basis for revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read arXiv:2412.13621. The genuinely new pieces are the bidirectional scandots (upward plus downward height scanning) and the centerline reward for keeping the robot on the pipe axis. The paper does well in simulation: Table IV is internally consistent and shows a clean curriculum effect — stage 1 fails on narrow pipes and obstacles, stage 2 handles narrow pipes but fails on obstacles, stage 3 reaches 93.75% on 0.2 m pipes with obstacles. The authors are also honest that the sim-to-real gap is large.\n\nThe soft spots are real, though. The central 'even with unexpected obstacles' claim rests on one row of Table V: Pipe C with obstacles, 46.5% 'Traversing Percentage', no trial count, no definition of the metric. If that is a success rate, the robot failed more than half the time. If it is the average fraction traversed, the robot never completed the pipe with obstacles. Either way, the obstacle-handling claim is not established. The real pipes are also 2–3 m long versus the 18 m training task, there are no baselines or ablations on the reward components, and no code or data are released. The 'first RL-based solution' claim is asserted without a literature search to back it. Minor but telling: the observation dimension is stated as R49 while the privileged vector alone is R296, so the notation is off.\n\nNone of this kills the core feasibility claim for clean pipes: the robot demonstrably traverses short PVC pipes of 0.19–0.24 m radius, which is nontrivial. But the obstacle-handling part of the abstract is not supported by the presented evidence, and the paper's own conclusion concedes degradation under noise and unseen obstacles.\n\nWho is this for? People working on legged locomotion in confined spaces and sim-to-real transfer for quadrupedal robots. It deserves serious refereeing because it proposes a new application with a plausible mechanism and internally consistent simulation results, but it needs major revision before acceptance: full trial counts, a clear metric definition, baselines, and ideally released code and data. I would send it to peer review with that expectation.","headline":"A plausible and partially demonstrated sim-to-real pipe-crossing result, with clean simulation ablations but thin real-world evidence that fails to support the obstacle-handling claim.","tokens_in":10379,"tokens_out":1493,"would_cite":false,"duration_ms":14918,"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":"The paper claims that reinforcement learning with bidirectional height scanning as privileged visual information, a centerline-based reward, and three-stage curriculum training can make a quadrupedal robot cross narrow pipes with…","keywords":["quadrupedal robot","narrow pipe inspection","reinforcement learning","sim-to-real transfer","privileged visual information","bidirectional scandots","curriculum learning","legged locomotion"],"falsifier":"Run the trained policy on a 10 m or longer pipe of radius 0.192 m over at least ten trials and count every success and failure, including the failures; the paper reports only selected successes, so a complete trial log that shows a success rate far below the simulated 93.75% would directly undermine the claim of adaptive pipe-crossing.","tokens_in":9226,"feed_emoji":"🤖","tokens_out":4207,"duration_ms":39032,"temperature":0.7,"pith_summary":"The paper tries to establish that a quadrupedal robot can be trained, entirely through reinforcement learning, to traverse narrow pipes that are too small for conventional inspection robots. It introduces bidirectional scandots—upward- and downward-facing height grids that give the policy privileged knowledge of the pipe floor and ceiling—and a reward function that keeps the robot near the pipe centerline while tracking velocity and avoiding collisions. Training proceeds in three stages that progressively narrow the pipe and add random obstacles, achieving a simulated success rate of 93.75% on a 0.2 m radius pipe with obstacles. On a real Unitree Go2 crossing PVC pipes of radius 0.242 m, 0.217 m, and 0.192 m, the policy completed traversals, although the authors note that real-world success was much lower than in simulation and that they report selected successful experiments. If correct, the work suggests a general-purpose legged robot could replace specialized wheeled, tracked, or inchworm pipe-inspection devices.","feed_headline":"RL-trained quadruped crosses narrow pipes with obstacles","feed_subtitle":"Two-way terrain scans and a centerline reward let a Unitree Go2 traverse 0.192 m radius pipes.","key_machinery":"The key machinery is bidirectional height scanning, called bidirectional scandots: an 11x12 grid of sample points around the robot that measures distances both downward to the pipe floor and upward to the pipe ceiling, producing a 296-dimensional privileged observation vector when combined with base velocity and environment parameters. This vector is encoded by a neural network during the first training stage, giving the policy simultaneous awareness of both surfaces that constrain motion inside a narrow pipe. The reward function's centerline-distance term, which penalizes deviation from the pipe's central axis, is what keeps the robot from scraping the walls, and the three-stage curriculum is what lets the policy first learn a stable gait before it is forced into tighter geometry and obstacle avoidance.","core_discovery":"The central claim is that a reinforcement-learned policy can adaptively cross narrow pipes when given a new form of privileged visual information: an 11x12 grid of height samples taken both below and above the robot, capturing the pipe floor and ceiling, which is encoded and used during training before being distilled into a policy that runs on depth-camera input alone. Combined with a three-stage curriculum (wide pipe, narrow pipe, narrow pipe with obstacles) and a reward that combines velocity tracking, energy penalties, collision avoidance, and a centerline-distance penalty, the policy reaches a 93.75% success rate over 16 trials in a simulated 0.2 m radius pipe with random obstacles. In the real world, the same policy crossed PVC pipes with radii of 0.242 m, 0.217 m, and 0.192 m, including a pipe with obstacles at a 46.5% traversing percentage, with the robot adjusting its gait and restoring balance despite slipping.","pith_inferences":["An implication left implicit is that the real-world statistics are based on selected successful trials rather than a complete trial count; a full account of failures would be needed to turn the demonstration into a deployment claim.","The paper's test pipes are only 2–3 m long, whereas the training task is 18 m; whether the gait and stability hold over longer distances real pipelines remains untested.","The reported 46.5% traversing percentage on the obstacle pipe suggests that the current policy is not yet dependable at the narrowest radius with obstructions, and combining the scandot encoder with a rapid-adaptation module could close the sim-to-real gap.","The bidirectional scan idea could be extended to full 360-degree rangefinding inside the pipe, which would supply the missing side-wall information the current grid only partially captures."],"forward_implications":["A standard quadruped with a depth camera could perform narrow-pipe inspection without custom wheeled or tracked mechanisms, lowering the cost of adapting robots to confined industrial environments.","The three-stage curriculum and bidirectional scan representation may transfer to other confined spaces with similar geometry, such as tunnels, ducts, or rubble voids.","The policy is stated to degrade under large visual noise or unseen obstacles, so practical deployment would need stronger perception robustness or LiDAR-based input.","Because the real-world success rate is reported to be much lower than in simulation, further work is needed on domain randomization or online adaptation before the method is reliable enough for routine inspection."],"supporting_citations":[{"why":"Provides the massively parallel Isaac Gym/Legged Gym training framework and the base reward terms for energy and collisions used throughout the method.","marker":"[28]"},{"why":"Supplies the heading-direction linear velocity reward and the depth-image policy input that the method adopts.","marker":"[26]"},{"why":"Demonstrates staged reinforcement learning for difficult terrain, which inspires the three-stage curriculum training.","marker":"[27]"},{"why":"Establishes a fully learned quadruped control baseline that the paper argues is unsuited to confined pipe environments.","marker":"[24]"},{"why":"Shows a simulated-only rapid adaptation method that the paper contrasts with its privileged-scan distillation approach.","marker":"[25]"}],"fun_headline_variants":["RL-trained robot dog masters narrow pipe crossing with obstacles","Quadruped robot learns to squeeze through tight pipes via RL","Unitree Go2 conquers narrow pipes down to 0.192 m radius","Reinforcement learning lets robot dogs traverse pipe obstacles","Narrow pipe inspection made possible by RL quadruped robot"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simulation's height-field pipe, with its randomized friction, mass, and bidirectional scans, is faithful enough to the real PVC pipe that a policy trained only in simulation keeps working on hardware.","fun_headline_variants_meta":{"raw":{"variants":["RL-trained robot dog masters narrow pipe crossing with obstacles","Quadruped robot learns to squeeze through tight pipes via RL","Unitree Go2 conquers narrow pipes down to 0.192 m radius","Reinforcement learning lets robot dogs traverse pipe obstacles","Narrow pipe inspection made possible by RL quadruped robot"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000398,"raw_usage":{"total_tokens":2035,"prompt_tokens":848,"completion_tokens":1187,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":464,"completion_tokens_details":{"reasoning_tokens":1102}},"tokens_in":464,"tokens_out":1187,"duration_ms":11319,"temperature":1.0,"reasoning_tokens":1102,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:57:09.207123+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained policy on a 10 m or longer pipe of radius 0.192 m over at least ten trials and count every success and failure, including the failures; the paper reports only selected successes, so a complete trial log that shows a success rate far below the simulated 93.75% would directly undermine the claim of adaptive pipe-crossing.","supporting_citations":[],"review_version":1}