{"id":"ab852b4a-3467-4e01-b9cd-348b1b72b435","arxiv_id":"2508.11275","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"A learned differentiable scalar field of reachability serves as a continuous constraint for humanoid motion planning.","lead":"This paper learns a smooth, differentiable reachability map for humanoid robots from sampled kinematic poses, then uses it as a constraint in continuous motion optimization. It targets faster, more flexible planning for footsteps, multi-contact motion, and loco-manipulation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Learned reachability map's boundary/gradient fidelity is unverified; without full-body feasibility checks, optimization solutions may be infeasible for the claimed humanoid tasks.","rationale":"The reader's verdict is UNVERDICTED because the full text is corrupted and cannot be technically audited. My stress-test agrees that the central claim's weakest point is the learned reachability map's fidelity: its positive region and boundary gradient must faithfully represent true reachability for the optimization to return valid humanoid plans. I do not see a demonstrable internal inconsistency from the abstract; rather, there is an unverified empirical precondition. The concern is load-bearing because the entire contribution—using a learned map as a constraint—depends on this fidelity. I also note a related but distinct issue: kinematic reachability alone may be insufficient for multi-contact and loco-manipulation tasks, which additionally require stability and collision-free whole-body configurations. However, since the full text is unavailable, I cannot determine whether the proposed optimization already includes such constraints; the abstract mentions the reachability map as a constraint but does not rule out additional terms. Thus, the appropriate verdict remains UNVERDICTED, not REJECT or ACCEPT. The proposed concrete test—full-body verification of optimizer outputs plus boundary accuracy metrics—would provide the missing evidence. I agree with the reader's identification of the weakest assumption; my extension to whole-body sufficiency does not change the central risk.","tokens_in":12841,"tokens_out":4395,"duration_ms":49114,"concrete_test":"Reproduce the paper's three planning experiments (footstep, multi-contact, loco-manipulation). For each returned plan, verify every contact/end-effector pose with a full-body forward-kinematics and static-dynamics solver that includes collision checking, joint limits, and center-of-mass stability. Record the fraction of plans that pass this complete verification. Separately, evaluate the learned map on a held-out dense grid of task-space samples generated uniformly in the workspace and labeled by the kinematic model: compute precision/recall of the positive region and, near the boundary, the cosine similarity between the learned gradient and the true boundary normal. If precision/recall is below 95% or gradient cosine similarity is below 0.9 in the boundary band, the claim that the map can be used directly as a constraint is unsupported. These checks would settle whether the central clai","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a differentiable reachability map, learned from end-effector poses sampled from a kinematic model, can be used directly as an optimization constraint for footstep, multi-contact, and loco-manipulation planning. This claim rests on two conditions: (i) the map's positive region accurately reflects the true set of kinematically reachable task-space poses, and (ii) the map's gradient near the boundary is sufficiently accurate so that the optimizer converges to feasible rather than spurious solutions. The abstract explicitly states the map is trained from kinematic-model samples; it does not state that the positive region is validated against full-body feasibility. For humanoid multi-contact planning, kinematic reachability of a foot or hand pose is necessary but not sufficient: the solution must also satisfy center-of-mass stability, joint limits, collision avoidance, and contact constraints. If the learned map is used in isolation, or if its kinematic-only positive region contains poses that are dynamically infeasible, the optimizer can return plans that fail the actual task. Moreover, the full manuscript text is corrupted (mojibake), so the training loss, network capacity, sample distribution, and any post-hoc boundary validation cannot be audited. The reader's weakest assumption correctly identifies the fidelity of the learned scalar field, but I extend it: even a perfectly learned kinematic reachability map may be insufficient for the claimed whole-body planning applications unless the optimization also enforces non-kinematic constraints. This is a correctness risk, not a claim about consensus, and it is directly testable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a differentiable reachability map: a scalar-valued task-space function, positive only in regions reachable by a robot end-effector, learned from sampled end-effector poses via a neural network or support vector machine. The map is then used as a differentiable constraint in continuous optimization for humanoid motion generation. The abstract claims efficient solutions to footstep planning, multi-contact motion planning, and loco-manipulation planning. However, the submitted full text is heavily corrupted (mojibake), so the technical content, equations, training details, and experimental results cannot be read or audited. The review is therefore based almost entirely on the abstract and the few readable fragments.","tokens_in":13188,"tokens_out":2949,"duration_ms":37609,"significance":"If the claims hold, the idea of replacing a binary/black-box reachability check with a smooth, differentiable scalar field is a useful contribution to humanoid motion planning. It would allow reachability constraints to be embedded directly into gradient-based optimization, potentially reducing computational cost and improving solution quality relative to sampling-based or derivative-free alternatives. The combination of a learned reachability model with continuous optimization for footstep, multi-contact, and loco-manipulation tasks is a plausible and relevant research direction. However, the current manuscript does not provide sufficient evidence to assess these claims: the full text is unreadable, and the abstract contains no quantitative comparisons, feasibility verification, or baseline results. The significance cannot be properly evaluated until a readable version with detailed experiments is provided.","major_comments":[{"comment":"The entire body of the manuscript is corrupted with mojibake (e.g., '�������� �������������� ...'). No equation, algorithm, or experiment can be checked. This is a blocking issue: the technical derivation, the learning formulation, the optimization setup, and the experimental evaluation are all inaccessible. The authors must provide a properly encoded PDF or source file.","section":"Full text"},{"comment":"The abstract claims that the proposed approach 'efficiently solves' footstep planning, multi-contact motion planning, and loco-manipulation planning, but no metrics, baselines, runtimes, success rates, or robot platforms are reported. Even for a short paper, quantitative evaluation is needed to support the efficiency and generality claims. Without such evidence, the central claim is unverified.","section":"Abstract"},{"comment":"The map is learned from end-effector poses generated using the robot's kinematic model. The abstract does not state whether the positive region of the learned map is validated against full-body feasibility. For humanoid multi-contact and loco-manipulation, kinematic reachability is necessary but not sufficient: solutions must also satisfy center-of-mass stability, joint limits, collision avoidance, and contact constraints. The paper should report what additional constraints are imposed and the rate at which optimized solutions pass a full-body feasibility check. If the reachability map is used alone, the optimizer may produce kinematically reachable but dynamically infeasible plans.","section":"Abstract and learning formulation"},{"comment":"The abstract mentions both neural network and SVM as learning models, but from the accessible fragments the loss function, network architecture, sample distribution, and regularization are unclear. Since the paper relies on the boundary and gradient accuracy of the learned map for optimization, the training procedure and validation strategy are load-bearing. The authors should specify how boundary fidelity is measured and whether the planner is tested on tasks or poses that are not part of the training samples.","section":"Learning setup (unreadable text)"}],"minor_comments":[{"comment":"The notation for task-space coordinates is not introduced in the abstract; it would help to state whether the map covers 3D positions only or 6D poses (position and orientation), as this affects the optimization formulation.","section":"Abstract"},{"comment":"Due to the corruption, related work and comparison to existing reachability representations (e.g., signed distance functions, occupancy maps, reachability databases) cannot be assessed. A readable manuscript should clearly position the contribution relative to these.","section":"Related work"},{"comment":"Any figures and tables in the full text are not legible. The authors should ensure that all figures are correctly embedded in the PDF and that captions are readable.","section":"Figures and tables"}],"recommendation":"uncertain","confidential_remarks":"The submitted PDF appears to be corrupted at the encoding level; the full text is unreadable mojibake. This may be a pipeline artifact, but as received, no substantive review of the technical content or experiments is possible. I recommend the editor ask the authors to resubmit a clean PDF. The abstract is suggestive of a useful idea, but the current submission cannot be evaluated for soundness or significance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: the full text is mojibake, so I'm judging from the abstract alone, same as the reader. The core idea is a sensible incremental contribution: replace discrete reachability grids with a learned scalar function that is positive in the reachable workspace and use it as a differentiable constraint in continuous optimization. That's not a new paradigm, but it's a useful upgrade for footstep, multi-contact, and loco-manipulation planning, and the paper is honest about the method—sampled end-effector poses from a kinematic model, a neural net or SVM surrogate, and three task types as demonstrations. No red flags in the abstract, no obvious circular fitting: the surrogate is trained on forward-kinematics samples, which is standard. Soft spots: we can't see training details, network capacity, sample coverage, baselines, or metrics. The efficiency claim is unquantified. The stress-test worry—that a kinematic reachability map alone cannot guarantee whole-body feasibility (CoM, collisions, contacts)—is a real question, but the abstract doesn't claim the map is the only constraint. More likely it's one constraint among others in the optimization, which would be standard practice. So that's a thing to check, not a confirmed flaw. The gradient near the boundary is also a practical risk: if the surrogate is optimistic near the reachable set, the optimizer could return kinematically infeasible poses. That's a common issue with learned constraints, and the paper may or may not address it with conservative margins or post-hoc verification. I can't tell. Bottom line: the idea is coherent and comes from a group that typically does careful work. The paper deserves a serious referee—not a desk reject. My own verdict stays provisional until I can read the actual text; no replication, no numbers, no audit trail. If you want to discuss it at reading group, I'd wait for a clean version, or treat it as a methods proposal and ask the obvious feasibility-validation question.","headline":"A clean, plausible extension of reachability maps—but the corrupted full text leaves soundness unverifiable; still deserves referee review.","tokens_in":592,"tokens_out":1208,"would_cite":false,"duration_ms":31548,"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 scalar reachability map learned from kinematic samples can serve directly as a continuous constraint, letting one optimization solve footstep, multi-contact, and loco-manipulation planning for humanoids.","keywords":["differentiable reachability map","humanoid motion generation","continuous optimization","footstep planning","multi-contact planning","loco-manipulation planning","neural network reachability","support vector machine"],"falsifier":"On a fixed robot model, densely sample forward-kinematics poses to train the map, then test it against a separate dense grid of task-space points checked by inverse kinematics. Count false positives (map says reachable, IK cannot find a pose) and false negatives (map says unreachable, IK succeeds). The claim would fail if optimization solutions that activate the reachability constraint are not reproducible as feasible IK solutions within tolerance.","tokens_in":12814,"feed_emoji":"🤖","tokens_out":6958,"duration_ms":71739,"temperature":0.7,"pith_summary":"This paper tries to establish that a robot's reachability—the set of positions a hand or foot can be placed in—can be compressed into one smooth scalar field, and that this field can be used directly as a constraint in numerical optimization instead of being checked point by point. The field is learned offline from end-effector poses sampled from the robot's kinematic model, using either a neural network or a support vector machine, and it is positive inside the reachable region and non-positive outside. Because the map is continuous and differentiable with respect to the end-effector position, planners can add it as an inequality constraint and let a gradient-based optimizer handle foot placement, contact choice, and manipulation simultaneously. The payoff of the claim is lower computation and a unified formulation for several hard humanoid planning problems, demonstrated on footstep planning, multi-contact motion planning, and loco-manipulation planning.","feed_headline":"A learned scalar field makes humanoid reachability a smooth constraint","feed_subtitle":"Once learned from pose samples, the same map drives footstep, multi-contact, and loco-manipulation planning.","key_machinery":"The central object is the differentiable reachability map: a scalar field $x \\mapsto f(x)$ over task-space coordinates, positive in the robot's reachable workspace and non-positive outside it, learned from forward-kinematic samples. Its role is to convert reachability from a set-membership test into an inequality constraint with a meaningful gradient; that is exactly what continuous optimization needs to push candidate footsteps or handholds into the reachable region while keeping all constraints smooth. The learning step absorbs the robot-specific geometry, so the same optimization formulation works for different robots and different tasks once the map is retrained.","core_discovery":"The paper's central claim is that the discrete predicate 'can this end-effector pose be reached?' can be replaced by a continuous scalar function $f(x)$ on task space, with $f(x)>0$ in reachable configurations and $f(x) \\le 0$ otherwise, such that the function and its gradient with respect to $x$ are available to an optimizer. It claims such a function can be learned from a database of end-effector poses generated by the robot's kinematic model, choosing either a neural network or a support vector machine as the learner. With this map, a humanoid motion generation problem becomes a constrained continuous optimization: candidate end-effector placements are variables, and $f(x) \\ge 0$ is one o","pith_inferences":["Beyond the paper: if the current map is position-only, a natural extension is to make it orientation-aware so 6D grasp poses are constrained directly rather than through a point approximation.","Beyond the paper: the reachability map could be composed with a whole-body kinematic Jacobian to expose how joint posture changes the reachable region, letting a single optimizer trade off body pose against end-effector placement.","Beyond the paper: a direct stress test of the representation is to train the map on one portion of the workspace and check whether its boundary generalizes; nonconvex holes in the reachable set are the likely failure mode for a globally smooth scalar field."],"forward_implications":["Footstep planning can be posed as one continuous optimization with reachability constraints, avoiding explicit combinatorial enumeration of candidate foot placements.","Multi-contact planning inherits the same reachability field, so support placements and contact transitions can be optimized in a unified cost-and-constraint formulation.","Loco-manipulation planning can include the same map as a constraint on the manipulated end-effector, coupling locomotion and manipulation in a single optimization.","Retraining the map for a new robot body or end-effector replaces hand-built reachability approximations with a data-driven function, as long as kinematic samples are available.","Because the map is differentiable, it can be reused as a penalty or cost term, not only as a hard constraint, in trajectory optimization."],"supporting_citations":[],"fun_headline_variants":["Smooth reachability maps make humanoid motion constraints differentiable","Learn a differentiable reachability map, then plan humanoid motions","Continuous reachability constraints for cheaper humanoid planning","From pose samples to smooth reachability constraints for humanoid motion","Differentiable reachability maps turn discrete checks into optimization gradients"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The map's positive region and its boundary gradient must faithfully match the true reachable workspace for the specific robot, which in turn depends on the sampled end-effector poses being representative of everything the kinematic model can actually reach.","fun_headline_variants_meta":{"raw":{"variants":["Smooth reachability maps make humanoid motion constraints differentiable","Learn a differentiable reachability map, then plan humanoid motions","Continuous reachability constraints for cheaper humanoid planning","From pose samples to smooth reachability constraints for humanoid motion","Differentiable reachability maps turn discrete checks into optimization gradients"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000808,"raw_usage":{"total_tokens":3363,"prompt_tokens":706,"completion_tokens":2657,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":450,"completion_tokens_details":{"reasoning_tokens":2576}},"tokens_in":450,"tokens_out":2657,"duration_ms":22280,"temperature":1.0,"reasoning_tokens":2576,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T19:59:56.323274+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a fixed robot model, densely sample forward-kinematics poses to train the map, then test it against a separate dense grid of task-space points checked by inverse kinematics. Count false positives (map says reachable, IK cannot find a pose) and false negatives (map says unreachable, IK succeeds). The claim would fail if optimization solutions that activate the reachability constraint are not reproducible as feasible IK solutions within tolerance.","supporting_citations":[],"review_version":1}