{"id":"637d8189-25bf-4ab8-b466-e14c57fc1a23","arxiv_id":"2505.23505","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A graph search over footsteps, grasps, and object poses, using switchable reachability maps, automatically plans humanoid loco-manipulation including rolling with regrasping.","lead":"Loco-manipulation planning for a humanoid robot is decomposed into object path planning, a footstep-and-grasp graph search, and whole-body motion generation, with reachability maps switched according to object motion. The framework is demonstrated in simulation on rolling a bobbin with regrasping, opening a door, and pushing a cart.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (7)'s two-sample movable-condition check is not shown to be conservative or sufficient; continuous CoM/object trajectories and rolling-angle transitions between reachability maps are unchecked.","rationale":"The reader's weakest assumption is the same as mine: Eq. (7) is the core transition-evaluation step, and if it admits false positives, the planner's output is not guaranteed to be executable. The paper lacks validation that would close this gap: no full-body WBM verification of the planned sequences (only Rviz animation and a ZMP check), no hardware experiment, and no quantitative comparison with [6]. I do not treat the absence of comparison as the central issue, because even a slower planner could be acceptable if correctness of the transition model were demonstrated. The proposed continuous replay test is feasible using the paper's own components (preview control, IK, reachability maps) and would directly determine whether the two-sample test is sufficient. If the test passes on many randomized transitions, the concern is resolved; if it fails, the movable condition must be strengthened. Therefore the reader's CONDITIONAL verdict is appropriate and unchanged.","tokens_in":11878,"tokens_out":8306,"duration_ms":77644,"concrete_test":"Replay the bobbin-rolling plan of Sec. VI-A1. For each transition that passed Eq. (7), reconstruct the actual CoM trajectory with the same preview-control law used in Sec. V and the object trajectory by interpolating along the Reeds-Shepp path with rolling-angle integration. Sample at 10 ms; for each sample, compute the instantaneous CoM pose and rolling distance and check whether the required grasping point lies in the corresponding reachability map (or solve full-body IK). Count transitions with violations strictly between t_mid and t_end. If any violation occurs, Eq. (7) is insufficient; a fix would be to require the whole sampled path, or a provably conservative inflated map, to be reachable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The framework's central claim rests on the movable-condition filter F_movable in Eq. (7) (Sec. IV-C3). This filter declares a footstep+object transition feasible if two discrete samples are reachable: the object's path midpoint with the CoM assumed at the fixed stance foot of the destination state, and the object's endpoint with the CoM at the mid-foot pose. No argument shows this is conservative or sufficient. Reachability maps are not convex in object pose or CoM position, so checking the 'most severe' CoM pose does not bound the other poses in the CoM trajectory unless monotonicity is proven. The actual CoM from preview control is not constrained to pass through the stance foot at the middle timing, and the object is not forced to move synchronously to the path midpoint; for rolling, the grasping point migrates with rolling angle and only finitely many maps (5 degree increments) are switched. Eq. (2f) even allows an index jump N_obj>1, so the midpoint tested is just the average of two path indices, not necessarily the geometric midpoint of the Reeds-Shepp curve. If any accepted transition becomes unreachable between the two sampled times, the planned footstep/regrasp sequence is not executable, undermining the claimed automatic rolling-with-regrasping results. The paper explicitly acknowledges the assumption but never validates it against a full WBM trajectory or experiment.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a three-stage planning framework for humanoid loco-manipulation: object path planning using RRT*, footstep and regrasping planning using AD* graph search with precomputed reachability maps, and whole-body motion planning using QP-based inverse kinematics. The main contribution is a transition model that represents combined locomotion and manipulation primitives, with feasibility evaluated by relocating and switching reachability maps according to the motion of the robot and the object. The framework is demonstrated in simulation on bobbin rolling with regrasping, door opening, and cart pushing tasks, with reported computation times of the order of seconds.","tokens_in":12164,"tokens_out":5596,"duration_ms":53889,"significance":"If the movability filter is reliable, the framework is a practically useful contribution to humanoid loco-manipulation planning: it is clearly formulated, modular, and computationally efficient, with initial solutions obtained in roughly 100 ms and the ability to handle rolling objects with automatic regrasping. The strengths include an explicit graph-search formulation, precomputed reachability maps that avoid repeated inverse kinematics, and a comparison of planning times against a prior method. However, the central claim rests on an assumption about CoM and object trajectories that is not validated against whole-body motion planning or hardware, which limits the strength of the feasibility claims. The paper also does not provide a formal correctness argument for the transition filter, so the current evidence is not sufficient to fully establish the purported versatility.","major_comments":[{"comment":"The movable condition F_movable is evaluated at only two discrete instants, under the assumption that the CoM is on the destination stance foot at the middle timing and at the mid-foot pose at the end, while the object moves synchronously to the path midpoint. No argument is provided that these two samples bound the hardest reachability condition over the continuous CoM and object trajectories. Reachability maps are not convex in CoM position or object pose, so checking the \"most severe\" CoM pose does not in general imply reachability for all intermediate CoM positions. The actual CoM from preview control (Section V) is not constrained to pass through the stance foot at the middle timing, and the object is not forced to move synchronously to the midpoint. For rolling objects, the reachability map is switched only at finite rolling-angle increments (5 degrees), leaving intermediate angles unchecked. Consequently, the planner may accept transitions that are not executable by the WBM layer. The paper explicitly acknowledges the assumption in Fig. 6(B), but neither proves its conservatism nor validates it against full whole-body trajectories or hardware experiments. This issue is load-bearing because it directly affects the feasibility of the planned footstep and regrasp sequences, including the claimed rolling-with-regrasping result.","section":"Section IV-C3, Eq. (7)"},{"comment":"The heuristic h_nominal is inadmissible, as the paper itself states (\"Without h_nominal, the heuristic is admissible\"). Since AD* relies on an admissible heuristic for its suboptimality guarantees, the reported \"heuristics inflation factor\" in Table II does not provide a rigorous bound on solution quality when h_nominal is used. The authors should either provide an analysis of the bound with h_nominal or explicitly state that the anytime optimality guarantee is heuristic only, so that readers are not misled about the formal properties of the planner.","section":"Section IV-B5, Eq. (5)"},{"comment":"The only dynamics validation of the planned motions is a ZMP check on a forward-rolling motion. ZMP staying within the support region does not guarantee that the hand can reach the object at all intermediate times, which is precisely what the movable condition (7) approximates. The paper should report whether the QP-based WBM planner actually finds feasible whole-body trajectories for the planned footstep and object sequences, and ideally execute the motions on a real robot, to substantiate the executability of the planned transitions. Without this validation, the experimental section is insufficient to support the claim that the rolling-with-regrasping motion is automatically planned and feasible.","section":"Section VI-B4 and Section V"}],"minor_comments":[{"comment":"The definition of c_mid_obj as c_P[(idx(c_obj[k]) + idx(c_obj[k+1]))/2] is ambiguous when the sum of the two indices is odd; please clarify whether the midpoint pose is obtained by interpolation, rounding, or some other rule.","section":"Section IV-C3, Eq. (7)"},{"comment":"The reachability map grid resolution is reported as 100 mm and 10 degrees in Section VI-B1, but the paper does not specify how an arbitrary object pose that does not coincide with a grid node is evaluated (e.g., nearest neighbor, bilinear interpolation). Please describe the evaluation rule.","section":"Section IV-C1"},{"comment":"The horizontal axis of Fig. 14(B) is labeled \"time [s]\", but the figure shows the length of the footstep sequence over time; please clarify whether this is computation time and what the unit refers to.","section":"Figure 14(B)"},{"comment":"The use of the Halton sequence to generate the footstep action set is interesting; please state the dimensionality of the sampled space and the number of samples used for the nominal action set, since these affect the planner's behavior.","section":"Section VI-B2"},{"comment":"The claim of being the first study to automatically plan rolling operation with regrasping could be tempered, since the experiments are simulations and prior work on non-prehensile manipulation may cover related tasks; a more careful comparison to existing rolling-manipulation literature would strengthen the novelty statement.","section":"Introduction, Contributions"}],"recommendation":"major_revision","confidential_remarks":"The paper is suitable in scope for a robotics letters venue, and the proposed framework is well structured. The main concern is that the movability filter in Eq. (7) is not proven conservative and is not validated against full whole-body execution, which is a central pillar of the claimed contribution. I would support publication after the authors either add a formal conservatism argument, refine the filter with more samples or interpolation, and provide WBM or hardware validation, or clearly frame the filter as a heuristic with its limitations discussed. The inadmissible heuristic issue in Section IV-B5 should also be addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's real contribution is the transition model: a graph-search state that couples stance/swing feet, hand labels, and object pose, with reachability maps relocated and switched as the object moves. That integration is new, and the rolling-with-regrasping result looks like a genuine first. The planning times are impressive — initial solutions in ~100 ms, final solutions in 1–2 s for fairly complex scenes — and the three demos (bobbin rolling, door opening, cart pushing) show the framework is versatile. The writing is clear and the equations are stated carefully. Credit where due: this is a solid, useful piece for the humanoid loco-manipulation subfield.\n\nThe weak spot is the movable-condition check in Eq. (7). It samples two instants — the middle and end of a footstep transition — under an assumed CoM trajectory and synchronous object motion. No argument shows this is conservative or sufficient; reachability maps are not convex, so checking the 'most severe' CoM pose does not bound intermediate poses without a monotonicity proof. For rolling objects, the maps are switched in 5-degree increments and the grasping point migrates, so the midpoint check is especially fragile. The paper explicitly acknowledges the assumption, which is honest, but it never validates it against a full WBM trajectory or a real robot. The one ZMP trace in Fig. 15 is suggestive, not systematic.\n\nAlso, the comparison with Jorgensen et al. is qualitative — they report their own times but don't run the same benchmarks on the prior method. And there's no sensitivity analysis for the grid size, step/regrasp costs, or the nominal foot pose heuristic. These are minor-to-moderate issues, not fatal ones. The framework is a planning heuristic; the assumption in Eq. (7) is a clear limitation that could be patched with post-hoc verification or a conservative construction.\n\nWho should read this: anyone working on humanoid loco-manipulation or footstep planning with manipulation constraints. It deserves a serious referee. A referee should ask for empirical validation of the movable condition on several random scenarios (comparing planned vs. full-WBM execution) and a direct timing comparison on shared benchmarks. The core idea holds up and the paper is worth engaging with.","headline":"Genuinely new integration of footstep and regrasp planning via switchable reachability maps, with a rolling-bobbin first, but validation is thinner than the claims.","tokens_in":12717,"tokens_out":2361,"would_cite":true,"duration_ms":24411,"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 graph search plans a humanoid's steps, grasps, and rolling as one problem.","keywords":["loco-manipulation planning","humanoid robots","graph search","reachability map","regrasping","object rolling","footstep planning","anytime dynamic A*"],"falsifier":"Take a planned bobbin-rolling sequence and execute it on a real or high-fidelity simulated humanoid while recording the actual CoM and object pose through each step; if the hand loses the grasp or the reachability condition fails at some intermediate time even though the middle and end poses passed, the transition model's most-severe-case assumption is false.","tokens_in":11695,"feed_emoji":"🤖","tokens_out":4578,"duration_ms":42327,"temperature":0.7,"pith_summary":"The paper proposes a planning framework that lets a humanoid robot decide, in one shot, where to step and which hand to grasp with while moving a large object. The central claim is that loco-manipulation can be written as a graph search over states that include both feet, the object pose, and the grasping hand, with transition feasibility checked by precomputed reachability maps. The maps are relocated and switched as the object rolls, so the moving grasp point is accounted for without running inverse kinematics during search. The authors demonstrate the planner on a bobbin-rolling task, a door-opening task, and a cart-pushing task, and report that it automatically generates a rolling-with-regrasping motion for the bobbin.","feed_headline":"A graph search plans a humanoid's steps, grasps, and rolling","feed_subtitle":"Switchable reachability maps let the planner fold rolling and regrasping into one search.","key_machinery":"The load-bearing mechanism is the transition model with switchable reachability maps. Its movable condition checks the object pose at the middle timing of a step against a map anchored at the stance foot, and the end pose against a map anchored at the midpoint of the two feet; for rolling objects, maps are switched according to the distance rolled since the last regrasp, so the changing grasp point is handled by map selection instead of inverse-kinematics calls during the search.","core_discovery":"The paper claims that a humanoid's loco-manipulation planning can be posed as a graph path search over states describing the stance foot, swing foot, object pose, and grasping hand, where each transition is accepted only if a reachability map certifies that the hand can hold the object at the middle and end of the footstep. For rolling objects, several reachability maps indexed by the distance rolled since the last regrasp are switched during search, capturing the motion of the grasp point without expensive kinematic recomputation. The authors report that this yields the first automatically planned motion of a large cylindrical object moved by rolling with regrasping.","pith_inferences":["The same switchable-map idea should extend to any object motion whose grasp point moves predictably, such as sliding, pivoting, or objects carried on a tray, without changing the search formulation.","Because the reachability maps are precomputed in posture space, the planner's runtime depends mostly on the graph search itself rather than on the complexity of the robot model.","The cost model's constant step and regrasp penalties could be replaced with task-specific costs such as energy, time, or stability margin without altering the transition model.","The assumption that checking only the middle and end poses guarantees reachability could be tested offline by sampling dense CoM and object trajectories from the generated whole-body motion, giving a quantitative safety margin for execution."],"forward_implications":["A single planner covers pure locomotion, pure manipulation, and combined stepping-while-moving-object behaviors, because the transition set includes no-stepping and zero-object-progress actions.","Regrasping emerges automatically when obstacles make it cheaper, since hand-label transitions are part of the search state, as shown in the door-opening case.","Rolling objects are handled in the same framework as sliding or fixed-grasp objects by switching reachability maps according to the distance rolled since the last regrasp.","An initial solution is found in roughly 100 milliseconds for the tested tasks, allowing the robot to keep moving while the anytime search improves the plan."],"supporting_citations":[{"why":"Provides the state-of-the-art loco-manipulation planning baseline and the learned reachability approach that this paper extends and compares against.","marker":"[6]"},{"why":"Introduces reachability-map-based planning for dual-arm regrasping, which the transition evaluation relocates and switches.","marker":"[12]"},{"why":"Supplies the RRT* sampling-based method used in object path planning for asymptotic optimality.","marker":"[21]"},{"why":"Provides Reeds-Shepp curves used to connect object poses when the object has car-like nonholonomic motion constraints.","marker":"[22]"},{"why":"Defines the Anytime Dynamic A* algorithm that gives the graph search its anytime and replanning properties.","marker":"[24]"},{"why":"Supplies the preview-control method that generates the CoM trajectory from the planned ZMP in whole-body motion planning.","marker":"[26]"}],"fun_headline_variants":["Graph search unifies humanoid walking, grasping, and rolling","Reachability maps guide humanoid loco-manipulation planning","One graph search for humanoid steps, grasps, and object rolling","Planning humanoid loco-manipulation: graph search with reachability maps","How humanoids plan to roll and regrasp: a graph search approach"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The feasibility check assumes the robot's center of mass passes through the stance foot at the middle of each step and the object moves to the middle pose exactly in sync, so checking only the middle and end poses is enough to guarantee reachability throughout the transition.","fun_headline_variants_meta":{"raw":{"variants":["Graph search unifies humanoid walking, grasping, and rolling","Reachability maps guide humanoid loco-manipulation planning","One graph search for humanoid steps, grasps, and object rolling","Planning humanoid loco-manipulation: graph search with reachability maps","How humanoids plan to roll and regrasp: a graph search approach"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000729,"raw_usage":{"total_tokens":3188,"prompt_tokens":792,"completion_tokens":2396,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":408,"completion_tokens_details":{"reasoning_tokens":2301}},"tokens_in":408,"tokens_out":2396,"duration_ms":18881,"temperature":1.0,"reasoning_tokens":2301,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:43:56.320808+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a planned bobbin-rolling sequence and execute it on a real or high-fidelity simulated humanoid while recording the actual CoM and object pose through each step; if the hand loses the grasp or the reachability condition fails at some intermediate time even though the middle and end poses passed, the transition model's most-severe-case assumption is false.","supporting_citations":[{"cited_title":"Finding locomanipulation plans quickly in the locomotion constrained manifold,","cited_arxiv_id":null,"evidence_quote":"Provides the state-of-the-art loco-manipulation planning baseline and the learned reachability approach that this paper extends and compares against."},{"cited_title":"Humanoid motion planning for dual-arm manipulation and re-grasping tasks,","cited_arxiv_id":null,"evidence_quote":"Introduces reachability-map-based planning for dual-arm regrasping, which the transition evaluation relocates and switches."},{"cited_title":"Optimal paths for a car that goes both forwards and backwards","cited_arxiv_id":null,"evidence_quote":"Provides Reeds-Shepp curves used to connect object poses when the object has car-like nonholonomic motion constraints."},{"cited_title":"Anytime dynamic A*: An anytime, replanning algorithm,","cited_arxiv_id":null,"evidence_quote":"Defines the Anytime Dynamic A* algorithm that gives the graph search its anytime and replanning properties."}],"review_version":1}