{"id":"17e21f71-1ba2-44e1-bd82-e0be41c004e9","arxiv_id":"2505.12934","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A legged robot repositions closely spaced rocks by deliberately triggering sand avalanches, using learned predictors of granular flow and of its own motion.","lead":"This robotics paper presents a learning-based system that lets a four-legged robot churn sand on a slope to slide rocks into desired positions while also moving itself toward a target. The authors report over 65 percent success in physical trials, well above a prior baseline.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing assumption is that EAA's midpoint action image and the manipulator-to-robot transfer keep the planned action aligned with actual leg-sand interaction; neither is directly validated.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing point: the EAA midpoint approximation and the manipulator-to-robot transfer of fe are unvalidated, yet they are the mechanism that connects the trained predictors to the physical deployment success rates. My independent reading of Sections 4.2, 4.3, and Appendix D confirms that these are not peripheral details—they determine whether the action image input to fe corresponds to the actual leg-sand interaction during robot locomotion. Without this alignment, the planned actions are optimized against the wrong environment predictions, so the reported 70% success could in principle arise from other mechanisms (notably direct leg-obstacle contact or the leniency of unspecified success criteria). The paper does provide meaningful independent support: 90 physical trials, ablations against GRAIN and other baselines, and a coherent decomposition of environment and robot state prediction. The concern is therefore not a refutation but a condition for accepting the central claim. The reader's CONDITIONAL verdict already captures this, and my analysis does not move it: the authors must validate or remove the EAA midpoint assumption and demonstrate that fe transfers to robot-captured images, ideally by releasing the recorded data and code so the offline replay can be run. I agree with the reader rather than adding a new objection; the trial-count inconsistency and lack of released artifacts, while real, are secondary to the scientific validity of the action-alignment assumption.","tokens_in":13930,"tokens_out":4146,"duration_ms":48760,"concrete_test":"Offline-replay the recorded deployment trials: reconstruct the action image under three variants—EAA midpoint x1, unadjusted x0, and endpoint x2—and feed each variant through fe, comparing predicted obstacle displacement to the actual observed displacement in the next depth image. If the midpoint variant does not reduce prediction MAE by more than its standard error relative to the no-EAA variant, the Section 4.3 assumption is unsupported. Separately, on robot trials where the robot is externally fixed, compare fe predictions on robot-captured depth images to ground truth; if the transfer MAE substantially exceeds the manipulator validation MAE, the Appendix D transfer assumption fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, 70% vs. 20% success on loco-manipulation, rests on a planning loop in which the environment predictor fe is fed an action image At. fe was trained only on gantry-manipulator data, yet deployment uses robot-captured depth images and actions adjusted by the EAA heuristic. Section 4.3 assumes the avalanche produced while the robot moves from x0 to x2 equals the avalanche that would be produced at the fixed midpoint x1=(x0+x2)/2; Appendix D further asserts, without direct evidence, that fe 'learns to ignore' manipulator/robot appearance differences. If either alignment is off, every planned action image is spatially misaligned with the actual leg-sand interaction region, and the receding-horizon cost minimization optimizes over systematically wrong predictions. The only quantitative evidence for EAA is Table 3, an MAE comparison on an experiment dataset showing obstacle-state MAE of 2.44 cm with EAA vs. 2.80 cm without it; this is not a deployment success-rate ablation. The transfer claim is supported only by the statement 'We believe the diffusion model learns to ignore these appearance differences.' This does not refute the empirical deployment results, but it leaves open the possibility that the reported success comes from contact-based pushing or task ease rather than the learned avalanche dynamics working as intended. The assumption is load-bearing because if it fails, the method's claimed novelty—planning through learned granular-flow prediction from a different platform—has not been demonstrated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DiffusiveGRAIN, a learning-based method for a quadrupedal robot to perform loco-manipulation on a granular sand slope by strategically inducing avalanches to reposition rocks. The system combines a U-Net-based diffusion environment predictor, a U-Net robot state predictor, and an 'Effective Action Adjustment' (EAA) heuristic to align the action representation with the predicted robot state. A receding-horizon planner with hand-tuned costs selects among six leg-excavation actions. Training uses 240 gantry-manipulator trials and 60 robot locomotion trials; deployment evaluates 90 physical trials. The central reported result is a 70% success rate on in-distribution loco-manipulation versus 20% for the GRAIN baseline, and 65–70% success in out-of-distribution trials with real rocks on 16°, 20°, and 24° slopes.","tokens_in":14205,"tokens_out":3630,"duration_ms":37909,"significance":"If the reported results hold, the paper makes a valuable empirical contribution: it demonstrates that a legged robot can leverage granular avalanche dynamics to indirectly manipulate multiple, closely spaced rocks while simultaneously moving itself, and that learned predictors can support such integrated planning. The authors provide substantial physical experimentation (375 systematic trials plus 90 deployment trials), multiple ablations (E2E-Diffusion, Pose-Diffusion, GRAIN, and DiffusiveGRAIN without EAA), and an out-of-distribution generalization test with real rocks. The idea of joint environment/robot state prediction for loco-manipulation is a plausible and interesting direction for terradynamics and legged robotics. However, the strength of evidence is limited by small per-condition trial counts, the absence of confidence intervals on success rates, and two unvalidated load-bearing assumptions (the EAA midpoint approximation and the manipulator-to-robot transfer of the environment predictor). The paper's significance is therefore conditional on additional validation of these assumptions and more statistically robust deployment measurements.","major_comments":[{"comment":"The EAA midpoint assumption, x1 = (x0 + x2)/2, is load-bearing for the entire planning loop: the adjusted action image A'_t is fed to fe, and if the assumed avalanche location is not the actual leg-sand interaction region, every planned step inherits a spatial misalignment. The only quantitative support in Table 3 is an obstacle-state MAE improvement from 2.80 cm (without EAA) to 2.44 cm (with EAA) on an experiment dataset, while robot-state MAE is identical (1.17 cm). This is not a deployment success-rate ablation. Please provide a deployment ablation (with vs. without EAA) and/or a direct measurement comparing the predicted avalanche centroid with the actual affected sand region.","section":"Sec. 4.3"},{"comment":"The environment predictor fe is trained exclusively on gantry-manipulator depth images but deployed on robot-captured depth images. Appendix D states 'We believe the diffusion model learns to ignore these appearance differences' as the only justification. This transfer is load-bearing for the central claim, because the planner optimizes over fe's predictions. Please add quantitative evidence: for example, compare fe's obstacle displacement predictions against ground truth on robot trials, or report a domain-shift analysis showing that the appearance difference does not systematically bias the predicted avalanche fields.","section":"Sec. 4.2 / Appendix D"},{"comment":"The headline loco-manipulation result, 70% vs. 20% success, is based on 10 trials per method. With 7/10 vs. 2/10 successes, the difference is not statistically significant at the 95% level (Fisher's exact test p ≈ 0.07). Success rates in Tables 1 and 2 are reported without confidence intervals, and the threshold that defines a successful trial is not explicitly stated. The phrase 'significantly high success rate' in the Conclusion is not supported by the reported statistics. Please report confidence intervals, exact p-values, or additional trials, and define the success criterion quantitatively.","section":"Sec. 5.3 / Table 1"},{"comment":"The receding-horizon planner depends on cost weights w1...w6 and parameters alpha, beta, gamma, all of which are hand-chosen. No sensitivity analysis is provided, so it is unclear whether the reported success is robust to these choices or whether the performance gap could change under plausible alternative weightings. Please add a sensitivity study, even on a subset of the deployment trials, to establish that the results are not an artifact of one specific tuning.","section":"Sec. 4.4 / 4.5"}],"minor_comments":[{"comment":"The title in the provided text reads 'Strategic Sand A valanche' with an erroneous space; please correct to 'Avalanche'.","section":"Abstract / Title"},{"comment":"The out-of-distribution results (65%, 70%, 65% for 16°, 20°, 24°) are based on 20 trials each, but no confidence intervals are given. Please add binomial confidence intervals or exact values for the success counts.","section":"Sec. 5.3 / Table 2"},{"comment":"The success criteria for manipulation, locomotion, and loco-manipulation trials are not explicitly defined in the main text (e.g., a distance threshold from the target). Please state them concretely, since repeatability and interpretation of the success rates depend on this definition.","section":"Sec. 4.5"},{"comment":"The figures rely on color cues (red horizontal line, green target) that may be ambiguous in grayscale. Consider adding labels or distinct markers to improve readability.","section":"Fig. 5 / Fig. 10"},{"comment":"There is a missing space in '0.3 mmglass beads'; please fix the typo.","section":"Sec. 2"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid empirical foundation but needs to address the two unvalidated assumptions (EAA and manipulator-to-robot transfer) with direct evidence, and to present the deployment success rates with proper statistical treatment. Given the modest sample sizes, the central claim as currently worded is stronger than the data support. These are fixable within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: DiffusiveGRAIN is a real step beyond GRAIN, and the paper deserves a proper review, but the headline numbers are thinner than they look and the key transfer assumption is asserted rather than tested.\n\nWhat's genuinely new: it removes two restrictions from the prior GRAIN line by using a diffusion model to predict multi-obstacle movement under coupled granular flow, and a separate U-Net to predict robot state changes from multi-leg excavation actions. The EAA trick—shifting the action image to the midpoint of the predicted robot motion—is a clever way to approximate the interaction region during a step. The hardware work is substantial: 90 deployment trials, ablations against E2E and pose-input baselines, and out-of-distribution trials with real rocks at three inclinations. Table 3 gives real evidence that the decomposition into fe and fr helps, and that EAA lowers obstacle-state MAE from 2.80 to 2.44 cm.\n\nThe soft spots are not fatal but they matter. The headline loco-manipulation result is 7/10 vs 2/10—twenty trials total. No confidence intervals anywhere. That's enough to show a trend, not a robust claim. The EAA midpoint assumption in Section 4.3 is exactly that: no direct measurement that the avalanche during a step matches the avalanche at the fixed midpoint. Table 3's MAE improvement is indirect evidence, but it doesn't close the loop on deployment. The bigger gap is manipulator-to-robot transfer. fe is trained on gantry-manipulator images and deployed on robot-captured depth images; Appendix D says \"We believe the diffusion model learns to ignore these appearance differences.\" That belief is doing real work, and I'd like to see it tested—for example, by comparing fe predictions against actual avalanche fields on the robot platform. Also hand-tuned weights (w1..w6) with no sensitivity analysis, no released code or data, and a minor trial-count inconsistency between the abstract and Section 5.1.\n\nNone of this sinks the paper. The contributions are clear and the experimental evidence is genuinely positive. Who should read it: roboticists working on legged locomotion, terrain adaptation, or granular manipulation. It's a better-than-average robotics paper that needs revision rather than rejection. I'd send it to review.\n\nMy verdict: conditional accept, with requests for confidence intervals, a direct transfer test, and artifact release.","headline":"DiffusiveGRAIN is a genuine empirical advance over GRAIN for rock repositioning on sand slopes, but the headline success rates rest on thin trial counts and the manipulator-to-robot transfer is more assumed than shown.","tokens_in":14759,"tokens_out":4618,"would_cite":true,"duration_ms":41202,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DiffusiveGRAIN claims that a quadruped can reposition closely spaced rocks on a sand slope by deliberately triggering avalanches, and that a diffusion-based environment and robot-state predictor lets it do so while moving to its own target.","keywords":["granular loco-manipulation","sand avalanche","diffusion model","legged robot","obstacle manipulation","granular media","robot state prediction","receding-horizon planning"],"falsifier":"Run the robot on a 20-degree slope with a single AF or RP action while a top-down depth camera records the avalanche; compare the measured depth-change field and obstacle displacements against the prediction made with the midpoint $x_1=(x_0+x_2)/2$. If the predicted avalanche footprint shifts from the measured one by more than roughly the obstacle radius, or the obstacle displacement error exceeds the paper's success threshold, the EAA premise is falsified.","tokens_in":13704,"feed_emoji":"🪨","tokens_out":11370,"duration_ms":105646,"temperature":0.7,"pith_summary":"DiffusiveGRAIN claims that a legged robot can reposition closely spaced rocks on a sand slope by deliberately triggering localized avalanches with its own legs, and that this manipulation should be planned jointly with the robot's locomotion. The paper argues that avalanche-driven rock motion is not independent: a nearby rock can block or redirect the flow, so a useful predictor must model several obstacles at once. It also argues that excavation actions change the robot's own position and orientation differently depending on which legs are used, so loco-manipulation needs a separate robot-state predictor. On physical trials, the method raised success on the joint loco-manipulation task from 20% for the prior GRAIN baseline to 70%, and reached 65-70% on out-of-distribution trials with real rocks at 16, 20, and 24 degree slopes. If these results hold, robots could actively reshape granular terrain to improve their own mobility.","feed_headline":"Avalanche planner lets a walking robot move rocks: 70% success","feed_subtitle":"A diffusion model predicts sand flow so a quadruped can shift closely packed rocks to targets on slopes up to 24 degrees.","key_machinery":"The machinery is two U-Net predictors over depth images plus a correction step. The environment predictor $f_e$ is a diffusion model that takes the current depth image $I_t$ and an RGB action image $A_t$ and outputs the sand-surface depth change $f_e(I_t,A_t)$; the robot-state predictor $f_r$ is a U-Net that outputs the robot's 2D position and 1D orientation change. Effective Action Adjustment (EAA) is the compensation step: it assumes the avalanche produced while the robot moves from $x_0$ to $x_2$ equals the avalanche a stationary manipulator would produce at the midpoint $x_1=(x_0+x_2)/2$, and it redraws the action image at $x_1$ before feeding it to $f_e$. Predicted robot pixels are then spliced into the predicted environment image, and a four-step, discounted receding-horizon search over six actions selects the sequence that minimizes a weighted cost of target distance and obstacle danger-zone proximity.","core_discovery":"On its own terms, the paper's discovery is that the granular flow that moves rocks and the robot state change caused by digging must be modeled together, and that a diffusion-based image predictor can do both if given an action representation and a midpoint correction. The environment predictor learns the depth-change field of the sand surface for one to five obstacles rather than treating each rock as an isolated object. The robot-state predictor learns the 2D position and orientation change caused by each of six leg-action groups. During execution, Effective Action Adjustment moves the action image to the midpoint between the robot's start and predicted end states, compensating for the fact that training data came from a fixed gantry rather than the moving robot. The claim is that this combined predictor, paired with four-step receding-horizon planning, is what lets a quadruped reach its own target while pushing rocks to theirs.","pith_inferences":["A direct extension would be to measure whether EAA's midpoint approximation degrades as excavation speed, step length, or slope angle increases; if so, a learned correction for the robot's continuous motion could replace the fixed midpoint.","The coupling result suggests a compact, distance-dependent interaction model, where rocks within a few centimeters suppress fore-aft displacement, could be combined with the learned image predictor to improve data efficiency.","Because the planner enumerates all six actions over a four-step horizon, scaling to larger action sets or longer horizons would likely require a learned policy or sampling-based planner rather than exhaustive search.","One testable consequence of the danger-zone cost is that a robot using this method should avoid direct leg-rock contact in most successful trials; counting contact events across trials would check whether the safety term is doing the work intended."],"forward_implications":["If the claim is right, a legged robot can treat granular slopes as partially shapeable terrain, using leg excavation both to move rocks and to reposition itself.","Planners for granular environments must model close obstacles as coupled; the paper reports that a second obstacle at 2 cm and 0 cm fore-aft distance cuts the first obstacle's displacement to 67% and 42%, respectively.","Leg-action groups give different locomotive and manipulative trade-offs: AF moves the robot fore-aft most, LP/RP rotate it most, and FP/LFE/RFE trigger localized avalanches with small robot state change.","The predictor transfers to unseen rock shapes and nearby slope angles without fine-tuning, with 65-70% success on out-of-distribution trials.","Removing EAA increases obstacle-state prediction error from 2.44 cm to 2.80 cm mean absolute error, suggesting the correction is load-bearing for accurate manipulation planning."],"supporting_citations":[{"why":"Prior learning-based granular manipulation method whose independence assumptions the paper tests and uses as the main baseline to beat.","marker":"[12]"},{"why":"U-Net architecture used as the backbone of both the environment and robot-state predictors.","marker":"[20]"},{"why":"Denoising diffusion probabilistic models, the training and sampling framework for the environment predictor.","marker":"[23]"},{"why":"Granular avalanche dynamics study motivating the choice of a 20-degree slope close to the angle of repose.","marker":"[15]"},{"why":"Avalanche dynamics on a rough inclined plane, cited to explain why adjacent obstacles affect avalanche propagation.","marker":"[33]"},{"why":"Prior result on leg-obstacle contact positions that motivates the danger-zone safety cost in planning.","marker":"[35]"},{"why":"Evidence that terrain surface state affects obstacle displacement under the same action, used to explain why pose-only predictors fail.","marker":"[36]"},{"why":"Vision Transformer architecture used by the baseline method, replaced here by U-Net-based diffusion predictors.","marker":"[32]"}],"fun_headline_variants":["Diffusion predicts sand flow and robot state to move rocks","Quadruped triggers strategic avalanches to reposition rocks","Joint sand-robot model lets robot shift rocks on slopes","Robot learns to use sand avalanches for rock relocation","Avalanche diffusion model moves rocks with over 65% success"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the avalanche generated by the moving robot's legs can be represented by the fixed-gantry manipulator's avalanche at the midpoint state $x_1=(x_0+x_2)/2$, and that the environment predictor trained on gantry data transfers to the robot without direct validation; if either fails, every planned step inherits a misaligned action.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion predicts sand flow and robot state to move rocks","Quadruped triggers strategic avalanches to reposition rocks","Joint sand-robot model lets robot shift rocks on slopes","Robot learns to use sand avalanches for rock relocation","Avalanche diffusion model moves rocks with over 65% success"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001985,"raw_usage":{"total_tokens":7747,"prompt_tokens":941,"completion_tokens":6806,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":6735}},"tokens_in":557,"tokens_out":6806,"duration_ms":51767,"temperature":1.0,"reasoning_tokens":6735,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:23:27.970730+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the robot on a 20-degree slope with a single AF or RP action while a top-down depth camera records the avalanche; compare the measured depth-change field and obstacle displacements against the prediction made with the midpoint $x_1=(x_0+x_2)/2$. If the predicted avalanche footprint shifts from the measured one by more than roughly the obstacle radius, or the obstacle displacement error exceeds the paper's success threshold, the EAA premise is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior learning-based granular manipulation method whose independence assumptions the paper tests and uses as the main baseline to beat."},{"cited_title":"Ronneberger, P","cited_arxiv_id":null,"evidence_quote":"U-Net architecture used as the backbone of both the environment and robot-state predictors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Denoising diffusion probabilistic models, the training and sampling framework for the environment predictor."},{"cited_title":"Gravish and D","cited_arxiv_id":null,"evidence_quote":"Granular avalanche dynamics study motivating the choice of a 20-degree slope close to the angle of repose."},{"cited_title":"B ¨orzs¨onyi, T","cited_arxiv_id":null,"evidence_quote":"Avalanche dynamics on a rough inclined plane, cited to explain why adjacent obstacles affect avalanche propagation."},{"cited_title":"Qian and D","cited_arxiv_id":null,"evidence_quote":"Prior result on leg-obstacle contact positions that motivates the danger-zone safety cost in planning."},{"cited_title":"loco-manipulation","cited_arxiv_id":null,"evidence_quote":"Evidence that terrain surface state affects obstacle displacement under the same action, used to explain why pose-only predictors fail."},{"cited_title":"Dosovitskiy, L","cited_arxiv_id":null,"evidence_quote":"Vision Transformer architecture used by the baseline method, replaced here by U-Net-based diffusion predictors."}],"review_version":1}