{"id":"11c2d78a-b395-48bf-a344-6dfc11140c58","arxiv_id":"2411.14386","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A perception-conditioned internal model lets humanoid robots climb 15 cm stairs and cross gaps with around 90% reported success, using onboard elevation maps.","lead":"This paper trains humanoid robots to walk over stairs and rough terrain by feeding them an elevation map built from onboard sensors, instead of raw camera images. The approach is fast to train and is tested on two different humanoids, which matters because robust perception is a current bottleneck for humanoid robots outside the lab.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 90% stair-climbing claim lacks trial counts and the policy's perception input is never validated against the online elevation map, so the central sim-to-real claim rests on an unmeasured assumption.","rationale":"The reader's conditional verdict is appropriate. My attack is a sharper version of the reader's weakest assumption: the online elevation map must match the ground-truth heights used in training, but the paper provides no error analysis for that map and no trial counts for the 90% staircase success rate. This is an empirical evidence gap rather than a conceptual flaw, so a stronger rejection is not warranted; however, the paper should add measured map error statistics and a defined number of repeated trials. Since this does not change the reader's conditional verdict, the verdict should remain unchanged.","tokens_in":8649,"tokens_out":4069,"duration_ms":44616,"concrete_test":"Run a fixed real-world protocol on H1: 10 consecutive 15 cm stair ascents with the stated LiDAR setup, logging per-run success/failure and recording the online elevation map heights at each step versus the known stair geometry. If the 10-run success rate is not ≥90% or the mean/max absolute map height error exceeds the tolerance set by the foot-length/stair-width constraint, the central claim is not supported. This single test settles both the perceptual-alignment assumption and the quantitative headline.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that PIM enables zero-shot deployment of humanoid policies on 15 cm continuous stairs with >90% success (Abstract, Section IV-B). That claim depends on the reconstructed elevation map matching the ground-truth height observations used in training: the policy is trained on exact obstacle heights (Section III-B, III-D), but inference uses heights sampled from an online map built from LiDAR/RGB-D plus odometry. Section III-D describes gravity alignment and ground filtering but gives no error, latency, drift, or noise analysis, and no comparison between map heights and ground-truth terrain. This matters acutely because the authors emphasize that the stair width equals the foot length, so a small height or placement error can cause the foot to catch the next step or lose balance. The same empirical gap affects the quantitative headline: Section IV-B asserts a 'success rate of over 90%' for continuous stair climbing but reports no number of trials, failures, or stopping criterion. The evidence provided is a narrative and selected images/videos. Thus the strongest possible validation—many consecutive successful ascents—and the error statistics that would confirm the perceptual assumption are absent. This is not an internal inconsistency; it is a load-bearing empirical assumption that is currently unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Perceptive Internal Model (PIM), a modification of the Hybrid Internal Model (HIM) in which the current elevation-map observation is concatenated with proprioceptive history for next-state estimation and is also fed to the policy. Policies are trained in simulation with ground-truth obstacle heights and deployed zero-shot on Unitree H1 and Fourier GR-1 using elevation maps built from a Mid-360 LiDAR or a RealSense T265+D435 setup. The paper claims continuous stair climbing with 15 cm steps at \"over 90%\" success, traversal of high platforms, gaps, and slopes, cross-platform generalization, and a training time of 3 hours on an RTX 4090 without depth-image rendering.","tokens_in":8897,"tokens_out":3431,"duration_ms":36858,"significance":"If the empirical claims were fully supported, the paper would be a useful step toward perceptive humanoid locomotion: it combines a proprioceptive internal model with a robot-centered elevation map, avoids depth-image rendering in training, and reports deployment on two different humanoid platforms. The design is simple and the reported training efficiency is attractive. However, the current manuscript is primarily a qualitative demonstration: the headline success rate lacks trial counts and definitions, the perception pipeline is not validated against ground truth, and the PIM-versus-HIM comparison is presented without numerical results. The paper's contribution would be strengthened considerably by quantitative hardware evaluation, map-error analysis, and seeded simulation comparisons; these are within the scope of the paper and, in my view, necessary before the central claims can be accepted.","major_comments":[{"comment":"The central quantitative claim, \"a success rate of over 90%\" for continuous stair climbing, is not supported by any reported trial count, failure definition, stopping criterion, per-run statistics, or sensor configuration. Because the paper uses this number as its headline result and compares against prior 10 cm stair results, the claim is not verifiable as written. Please add a hardware evaluation table with the number of attempts, success criteria, and confidence intervals for each robot and sensor setup.","section":"Abstract and Section IV-B"},{"comment":"The training-to-inference perception gap is load-bearing but unmeasured. Policies are trained on exact ground-truth obstacle heights (Sections III-B and III-D), while inference uses heights sampled from an online elevation map reconstructed from LiDAR or RGB-D plus odometry. The manuscript describes gravity alignment and ground-point filtering but gives no error, latency, drift, or noise analysis and no comparison between reconstructed map heights and ground-truth terrain. Given the stated constraint that the stair width equals the foot length, small height or placement errors can plausibly cause failure. Please provide quantitative map-error evaluation (for example, height RMSE against ground truth under typical robot motion) and, if possible, a simulation study that perturbs the sampled heights to demonstrate policy robustness to the expected perception noise.","section":"Section III-D and Section IV-B"},{"comment":"The PIM-versus-HIM comparison is only qualitative. Fig. 5 reports \"estimation loss\" and \"terrain level\" without axis labels, units, seeds, or numerical values, so the statement that PIM \"achieves higher training efficiency and higher estimation accuracy\" is not established. Please report quantitative velocity-estimation error, terrain-level progression over training, and final task success rates over multiple seeds, with error bars or confidence intervals.","section":"Section IV-A and Fig. 5"},{"comment":"The cross-platform claim for the Fourier GR-1 is supported only by narrative descriptions and images. No task success rates, number of attempts, failure counts, or sensor specifications are given for GR-1 on stairs or other terrains. Please add per-platform quantitative results, including the sensor configuration used in each deployment, so that the claimed generality across robot platforms can be assessed.","section":"Section IV-D"},{"comment":"The claim of surpassing prior methods that achieve only 10 cm stairs is not supported by a direct comparison under matched conditions. Different robot hardware, training protocols, and evaluation procedures confound the comparison. Please either run baseline policies on the same robot and terrain, or soften the claim to a statement about demonstrated capability rather than comparative superiority.","section":"Section IV-B"}],"minor_comments":[{"comment":"The displayed equation for the symmetry loss has unbalanced parentheses and an unclear argument structure; please rewrite it with explicit closing parentheses and define each symmetry operator's output space.","section":"Equation (1), Section III-F"},{"comment":"Several reward entries are malformed or undefined, in particular the \"Joint power\" equation, the \"Feet slip\" term containing \"~1new contact\", and the \"Joint tracking error\" term with an ambiguous subscript. Please provide corrected equations and define all indicator functions and constants.","section":"Table I"},{"comment":"The sampling procedure for the 96 elevation points is underspecified: no grid resolution, coordinate convention, or ordering is given. Please clarify how the 0.8 m x 1.2 m region is discretized and how the sampled points are aligned with the gravity-aligned map frame. Also, \"Terrian\" in the Fig. 3 caption is a typo.","section":"Section III-D and Fig. 3"},{"comment":"The axes of Fig. 5 are not labeled and the units are unclear; please label both axes and define \"estimation loss\" and \"terrain level\" in the caption or text.","section":"Fig. 5"},{"comment":"The manuscript contains several grammatical errors and unconventional spellings (for example, \"we concluded as PIM\", \"Realsense\" instead of \"RealSense\", and \"V ar\" without a subscript). A careful proofreading pass would improve readability.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The main issue is not novelty or circularity: PIM is legitimately built on HIM as a component, and the paper does not reduce its result to a fitted parameter. The concern is an evidence deficit in the quantitative hardware claims, which is fixable within the manuscript's scope by adding trial counts, map-error validation, and seeded numerical comparisons. I would not reject on the current evidence, but I would require the quantitative material before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a plausible, well-scoped extension of the authors' own HIM to perceptive humanoid locomotion. The genuinely new piece is the integration of online elevation maps into the state predictor, not any single component: elevation maps for quadrupeds, PPO, and symmetry regularization are all known. What the paper adds is a specific way to feed terrain height samples into HIM's next-state estimator, avoiding depth rendering and keeping training cheap (3 hours on one GPU).\n\nThe empirical work is real and goes beyond the current literature: two different humanoids (H1 and GR-1), 15 cm stairs versus the 10 cm reported for prior methods, and zero-shot deployment with LiDAR and RGB-D setups. That is a useful engineering result worth taking seriously. The authors also clearly state the hard constraint—stair width equals foot length—so the task is genuinely perceptually demanding.\n\nThe soft spots are exactly where the stress-test note lands. The >90% continuous stair-climbing success rate appears in the abstract and results but comes with no trial count, no failure count, no stopping criterion, and no variance. That is a headline number with no measurement behind it. The second, more load-bearing gap is the elevation map. The policy is trained on exact ground-truth heights, but inference uses heights sampled from an online map built from LiDAR/RGB-D plus odometry. Section III-D describes gravity alignment and ground filtering, but gives no error analysis, no latency or drift characterization, and no comparison between mapped heights and surveyed terrain. Given that a small height or placement error can catch a foot on the next step, this unvalidated perceptual assumption is the core sim-to-real risk.\n\nThe PIM-versus-HIM comparison in Fig. 5 is qualitative and only shows training curves, not hardware. No code or data are released, so independent confirmation is not possible. These are empirical gaps, not conceptual flaws. The circularity burden is low: HIM is a building block, not the conclusion. Nothing here is internally inconsistent.\n\nWho gets value from this? Researchers working on humanoid sim-to-real will find the integration idea useful and the cross-platform demos encouraging, but they should not treat the quantitative claims as established. A serious editor should send this to peer review—the method is timely and the results are plausible—but the review should explicitly demand trial counts, map error statistics, and ideally code release before the claims can be accepted.","headline":"Perceptive extension of HIM that plausibly climbs 15 cm stairs on two humanoids, but the headline 90% success and online-map accuracy are unmeasured.","tokens_in":9448,"tokens_out":1430,"would_cite":false,"duration_ms":16381,"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 perception-augmented state predictor lets humanoid robots climb 15 cm stairs with over 90% success on two platforms.","keywords":["humanoid locomotion","perceptive control","elevation map","sim-to-real transfer","hybrid internal model","stair climbing","reinforcement learning","zero-shot deployment"],"falsifier":"Feed the policy an elevation map whose heights are corrupted by a known bias or latency, for example all step heights reduced by 5 cm or delayed by 200 ms, and measure the stair-climbing success rate; if the rate collapses or the robot stops lifting its feet, it would show that the reported performance depends on near-exact map heights rather than on the learned state predictor's robustness.","tokens_in":8447,"feed_emoji":"🤖","tokens_out":6295,"duration_ms":55292,"temperature":0.7,"pith_summary":"This paper argues that a humanoid locomotion policy can be made reliable on uneven ground by feeding the controller a continuously updated elevation map of the terrain around the robot, not as a vision input but as part of the robot's internal state predictor. The proposed Perceptive Internal Model (PIM) takes ground-truth obstacle heights during training and sampled heights from a LiDAR or RGB-D elevation map at deployment, extending the Hybrid Internal Model's next-state estimation. The authors report zero-shot sim-to-real transfer to two different humanoids, with over 90% success climbing 15 cm stairs and successful traversal of gaps, slopes, and high platforms. If true, this would be the first demonstration of a single perceptive humanoid locomotion policy that transfers across robot platforms, and it would show that perceptive information can improve state estimation without extra training stages or simulated depth rendering.","feed_headline":"Terrain-sensing predictor lets humanoids climb 15 cm stairs","feed_subtitle":"Two robots deploy zero-shot after a three-hour training run, with over 90% stair success.","key_machinery":"The key mechanism is the Perceptive Internal Model (PIM): a state predictor that ingests both the recent proprioceptive history and the current perceptive observation, a 96-point sampling of heights in a 0.8 m by 1.2 m grid around the robot's base link aligned to gravity. It produces a velocity estimate trained by regression to simulator ground truth and a latent next-proprioception variable trained by contrastive learning, and both feed the policy along with the current observation. During training the policy sees exact terrain heights; during deployment those heights come from a continuously updated elevation map built from point clouds and odometry, so the robot effectively perceives the terrain beneath its feet while remaining less affected by camera motion and noise than raw depth inputs.","core_discovery":"The central claim is that augmenting the learned state predictor with current terrain elevation leads to more accurate next-state estimates and therefore to locomotion policies that can handle fine-grained foothold tasks such as continuous 15 cm stair climbing. PIM builds on HIM, which estimates the robot's next linear velocity by regression and its next proprioceptive state by contrastive learning; PIM concatenates the current sampled elevation map with the proprioceptive history so that the predictor can use the ground under the feet. In simulation the policy is trained on exact obstacle heights, and at inference the same observation is filled from an onboard elevation map built from LiDAR or RGB-D data in a gravity-aligned map frame. The authors report that this single-stage training completes in about three hours on one RTX 4090 and deploys without fine-tuning on Unitree H1 and Fourier GR-1, including continuous stair climbing with success rate above 90% and jumps onto platforms and over gaps.","pith_inferences":["If the mechanism is as general as the cross-platform result suggests, the same elevation-map-augmented state predictor could be applied to other bipedal or quadrupedal platforms, but that transfer is an extrapolation beyond the paper's two robots.","The reported sensitivity to foot-length versus stair-width suggests that foothold precision, not raw perception range, is the limiting factor; a testable corollary is that adding a foothold-planning head would push success further on taller stairs.","The lack of reported map-error analysis means the practical ceiling of the method may be set by the elevation-map pipeline rather than by the learned policy; injecting calibrated map noise in simulation would quantify that ceiling.","A natural extension is to let the policy also learn from uncertain or partial elevation maps during training, which would make the sim-to-real gap explicit instead of relying on the map being exact."],"forward_implications":["Continuous 15 cm stair climbing at over 90% success becomes feasible, a capability the paper reports earlier methods cap at 10 cm.","The same training pipeline transfers zero-shot to two physically different humanoids, suggesting the method is not tied to one robot's mass, height, joint configuration, or foot design.","Because no depth-image rendering is needed in simulation, policy training runs in about three hours on a single RTX 4090 without a second training stage.","Using the elevation map inside the state predictor, rather than only as a visual input, lowers next-state estimation loss and lets the policy handle harder terrain than the same policy with a non-perceptive internal model.","The policy produces coordinated whole-body movements, including arm swings during jumps, without imitation data or prior motion trajectories."],"supporting_citations":[{"why":"Supplies the Hybrid Internal Model that PIM extends with perceptive observations for next-state estimation.","marker":"[3]"},{"why":"Provides the GPU elevation mapping module used to build the onboard height map from point clouds and odometry.","marker":"[27]"},{"why":"Perceptive humanoid baseline that reportedly cannot handle continuous stairs, serving as the comparison PIM aims to beat.","marker":"[10]"},{"why":"Prior humanoid method reported to reach only 10 cm stairs, the baseline for the stair-climbing claim.","marker":"[6]"},{"why":"Prior method also capped at 10 cm stairs, another baseline for the claimed 15 cm capability.","marker":"[8]"},{"why":"Proximal policy optimization, the RL algorithm used to train the policy and value networks.","marker":"[28]"},{"why":"FAST-LIO LiDAR-inertial odometry used with the Mid-360 LiDAR to provide odometry for the elevation map.","marker":"[30]"},{"why":"FAST-LIO2, the follow-up odometry package used for the LiDAR-based elevation map setup.","marker":"[31]"},{"why":"Supplies the symmetry regularization losses used to improve gait symmetry.","marker":"[32]"}],"fun_headline_variants":["Terrain-aware internal model lets humanoids climb stairs in 3 hours","3-hour training enables humanoid stair climbing with 90% success","PIM: 3-hour training, 90% stair climbing success on humanoids","Elevation maps in internal model: 3-hour training for 90% stair success","Humanoid learns stair climbing in 3 hours with elevation-aware predictor"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the online elevation map reconstructed from LiDAR or RGB-D data and odometry reproduces the ground-truth obstacle heights used in simulation closely enough for zero-shot transfer, since the paper reports no quantitative error, latency, or drift analysis for the map.","fun_headline_variants_meta":{"raw":{"variants":["Terrain-aware internal model lets humanoids climb stairs in 3 hours","3-hour training enables humanoid stair climbing with 90% success","PIM: 3-hour training, 90% stair climbing success on humanoids","Elevation maps in internal model: 3-hour training for 90% stair success","Humanoid learns stair climbing in 3 hours with elevation-aware predictor"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001348,"raw_usage":{"total_tokens":5487,"prompt_tokens":968,"completion_tokens":4519,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":4419}},"tokens_in":584,"tokens_out":4519,"duration_ms":29478,"temperature":1.0,"reasoning_tokens":4419,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:13:30.295532+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed the policy an elevation map whose heights are corrupted by a known bias or latency, for example all step heights reduced by 5 cm or delayed by 200 ms, and measure the stair-climbing success rate; if the rate collapses or the robot stops lifting its feet, it would show that the reported performance depends on near-exact map heights rather than on the learned state predictor's robustness.","supporting_citations":[{"cited_title":"Hybrid internal model: Learning agile legged locomotion with simulated robot response,","cited_arxiv_id":null,"evidence_quote":"Supplies the Hybrid Internal Model that PIM extends with perceptive observations for next-state estimation."}],"review_version":1}