{"id":"ebc81691-eee1-4ca9-b475-5206e35f29e3","arxiv_id":"2605.30778","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Hierarchical object-informed MPPI control raises success rates in obstacle-avoiding object pushing by 40% in simulation and 20% on hardware compared to standard MPPI.","lead":"The paper proposes a hierarchical model predictive path integral control approach that first solves an object-only planning problem and then uses the resulting trajectory as a reference for joint robot-object planning in non-prehensile pushing tasks. A smart generalist might read it because improved long-horizon planning for pushing objects around obstacles could make warehouse and service robots more practical.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Object-only direct-actuation trajectory may be dynamically infeasible for the robot, weakening its value as reference signal","rationale":"The reader's weakest_assumption directly identifies the same load-bearing step; the concrete ablation above would falsify or confirm whether that step actually drives the measured gains. Full-text access does not remove the assumption, so the verdict moves from UNVERDICTED to CONDITIONAL pending the check.","tokens_in":1643,"tokens_out":315,"duration_ms":11884,"concrete_test":"Re-run the simulation benchmark suite with the reference trajectory replaced by a straight-line interpolation from initial to goal pose (same MPPI hyperparameters and sampling budget); if task success rate drops by more than 15 percentage points relative to the published object-informed numbers, the reference signal is load-bearing for the headline claim.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The method's central construction (hierarchical MPPI) requires that the separately solved object trajectory—computed under the assumption of direct object actuation—supplies an accurate and useful reference for the subsequent joint robot-object MPPI. Because the object-only solve ignores robot kinematics, reachability, and contact force limits, the reference can lie outside the feasible set inducible by the 6-DoF arm; in such cases the reference term in the joint cost can bias sampling toward unreachable states rather than improve long-horizon planning. The reported 40 % / 20 % success gains rest on this reference being net-positive across the tested pushing tasks with obstacles.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes a hierarchical MPPI controller for non-prehensile pushing tasks. An object-only trajectory is first computed under the assumption of direct actuation; this trajectory is then inserted as a reference term into a joint robot-object MPPI optimization. The authors report that the resulting controller raises task success by 40 % (simulation) and 20 % (hardware) relative to standard MPPI while maintaining or improving control frequency on a 6-DoF xArm6 manipulator performing obstacle-avoidance pushes.","tokens_in":1770,"tokens_out":446,"duration_ms":29084,"significance":"If the performance claims are shown to be robust, the hierarchical reference-signal construction offers a lightweight way to inject long-horizon object-level information into sampling-based MPC without expanding the full state space, which could be useful for other underactuated contact-rich tasks.","major_comments":[{"comment":"§3 (Hierarchical formulation): The central claim that the object-only reference improves joint planning rests on the unexamined assumption that a trajectory generated under direct object actuation remains a net-positive guide once robot kinematics, reachability, and contact-force limits are enforced. No analysis or counter-example is supplied showing when this reference lies outside the robot-inducible set; if such cases occur, the added cost term can bias sampling toward unreachable states and undermine the reported 40 % / 20 % success gains.","section":"§3"},{"comment":"§5 (Experimental evaluation): The quantitative results (40 % success increase, 26 % faster frequency in simulation; 20 % in hardware) are presented without reported trial counts, standard deviations, statistical tests, or explicit baseline definitions. Because these numbers are the primary evidence for the method’s advantage, the absence of this information prevents verification that the data support the central claim.","section":"§5"}],"minor_comments":[{"comment":"Notation for the reference cost term and the weighting between object-only and joint stages should be introduced once and used consistently.","section":"§3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback on our manuscript. We address each major comment below and will incorporate revisions to improve clarity and rigor.","responses":[{"response":"The object-only trajectory is generated using the same object dynamics and obstacle-avoidance costs as the joint planner, differing only in the direct-actuation assumption. The reference enters the joint MPPI cost as a soft quadratic penalty rather than a hard constraint, permitting sampled trajectories to deviate when robot reachability or contact limits render the reference unreachable. In the evaluated pushing tasks, the object-only plans remained within the inducible set because the robot maintains continuous contact with the object. We agree, however, that the manuscript would benefit from explicit discussion of this point. We will add a short analysis subsection in §3 examining the conditions under which the reference remains beneficial and noting potential degradation when the reference lies far outside the reachable set.","revision_made":"yes","referee_comment":"[§3] §3 (Hierarchical formulation): The central claim that the object-only reference improves joint planning rests on the unexamined assumption that a trajectory generated under direct object actuation remains a net-positive guide once robot kinematics, reachability, and contact-force limits are enforced. No analysis or counter-example is supplied showing when this reference lies outside the robot-inducible set; if such cases occur, the added cost term can bias sampling toward unreachable states and undermine the reported 40 % / 20 % success gains."},{"response":"We acknowledge that the current manuscript omits the requested statistical details. The reported gains are computed from repeated trials whose exact counts, standard deviations, and significance tests are not stated. We will revise §5 to report the number of trials performed for each method, the observed standard deviations, the explicit definition of the baseline (standard MPPI without the object reference term), and the results of appropriate statistical tests. These additions will allow readers to verify the strength of the performance claims.","revision_made":"yes","referee_comment":"[§5] §5 (Experimental evaluation): The quantitative results (40 % success increase, 26 % faster frequency in simulation; 20 % in hardware) are presented without reported trial counts, standard deviations, statistical tests, or explicit baseline definitions. Because these numbers are the primary evidence for the method’s advantage, the absence of this information prevents verification that the data support the central claim."}],"tokens_in":1332,"tokens_out":514,"duration_ms":20609,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper splits MPPI into an object-only stage that assumes direct actuation, then feeds the resulting trajectory as a reference into the joint robot-object optimization. This is the concrete change they make to standard MPPI for long-horizon non-prehensile pushing.\n\nWhat the work actually shows is that the reference helps on the tested tasks. In simulation they report 40% higher success and 26% faster control rate; on the xArm6 hardware they see a 20% success lift with similar compute. The tasks involve reaching a goal while avoiding obstacles, so the non-myopic aspect matters.\n\nThe formulation itself is straightforward once you see the cost terms. They keep the usual MPPI sampling but add a term that penalizes deviation from the precomputed object path. That is the part that is new relative to plain MPPI.\n\nThe soft spot is exactly the one flagged in the stress-test note. Because the object-only solve ignores robot kinematics and contact limits, the reference can sit outside the set of states the arm can actually produce. When that happens the extra term can steer samples toward unreachable states rather than useful ones. The abstract gives no ablations on reference weight or failure cases where the object plan is badly mismatched, so it is not yet clear how often this hurts versus helps. If the full paper only shows the cases where it works, the 40% and 20% numbers are harder to generalize.\n\nThe experiments are on hardware and the method is simple enough to reimplement, which is the main practical value. Readers who already use MPPI for contact-rich manipulation will want the details on how the reference is injected and what the compute overhead looks like. It is not a foundational result, but it is a usable incremental technique.\n\nI would bring this to a reading group for the implementation choices and the real-robot numbers. It deserves peer review because the central idea is testable, the claims are quantified on hardware, and the potential flaw is narrow enough that referees can check it directly.","headline":"The hierarchical MPPI with an object-level reference trajectory delivers clear success gains on pushing tasks, but its value depends on how often that reference stays reachable for the arm.","tokens_in":2261,"tokens_out":488,"would_cite":false,"duration_ms":19169,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Hierarchical MPPI first plans an object trajectory assuming direct actuation then uses it as reference for joint robot-object optimization.","keywords":["non-prehensile manipulation","model predictive path integral","hierarchical control","object pushing","robot manipulation","long-horizon planning","MPPI"],"falsifier":"Running identical pushing tasks with and without the object reference and observing no improvement or a drop in success rate when the reference is supplied.","tokens_in":2568,"feed_emoji":"🤖","tokens_out":645,"duration_ms":17345,"temperature":0.7,"pith_summary":"The paper establishes that solving a simplified object-only problem first and feeding its trajectory as a reference into the full MPPI optimization improves success on long-horizon non-prehensile tasks such as pushing an object to a goal while avoiding obstacles. Standard MPPI struggles with the computational cost of predicting underactuated and discontinuous contacts over many steps; the two-stage split keeps the expensive joint optimization tractable while still incorporating object-level foresight. In simulation the method raises success by 40 percent at 26 percent higher control frequency, and on hardware it raises success by 20 percent at comparable runtime. A sympathetic reader would therefore view the approach as a practical way to extend model-predictive control to manipulation problems that require non-myopic reasoning without simply scaling up compute.","feed_headline":"Object reference in MPPI raises pushing-task success 40%","feed_subtitle":"Precomputed object trajectory under direct actuation guides joint planning for higher success at similar or lower cost.","key_machinery":"The hierarchical MPPI that separates object-only planning (direct actuation assumed) from joint robot-object planning and uses the object trajectory as reference.","core_discovery":"The central claim is that a hierarchical MPPI formulation, in which an object-only plan computed under direct actuation serves as a reference signal for the subsequent joint robot-object MPPI, produces higher task success than standard MPPI on non-prehensile pushing tasks that require long-horizon prediction.","pith_inferences":["The same two-stage split may reduce the impact of model mismatch in the robot dynamics during the first planning stage.","If the object reference is occasionally inaccurate the joint optimizer can still correct it, suggesting the method is robust to moderate errors in the simplified model.","The approach could be tested on tasks with movable obstacles or multiple objects by extending the object-only stage accordingly."],"forward_implications":["Task success rises 40 percent in simulation while control frequency increases 26 percent.","Task success rises 20 percent on hardware experiments with similar computation time.","Long-horizon prediction for underactuated discontinuous contacts becomes feasible without increasing the planning horizon length.","The same separation applies to other object-pushing scenarios that involve static obstacles."],"fun_headline_variants":["Object reference MPPI raises push success by 40 percent","MPPI hierarchy with object plan raises task success 40%","Object-only plan references MPPI for higher push success","Object-informed MPPI control raises non-prehensile success"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"That the object trajectory obtained by assuming direct actuation remains a sufficiently accurate and useful reference once the robot's actual limited actuation is considered in the joint optimization.","fun_headline_variants_meta":{"raw":{"variants":["Object reference MPPI raises push success by 40 percent","MPPI hierarchy with object plan raises task success 40%","Object-only plan references MPPI for higher push success","Object-informed MPPI control raises non-prehensile success"]},"model":"grok-4.3","cost_usd":0.006452,"raw_usage":{"total_tokens":2980,"prompt_tokens":583,"num_sources_used":0,"completion_tokens":65,"cost_in_usd_ticks":64524500,"prompt_tokens_details":{"text_tokens":583,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2332,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":583,"tokens_out":65,"duration_ms":16250,"temperature":1.0,"reasoning_tokens":2332,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-28T22:31:36.428612+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Running identical pushing tasks with and without the object reference and observing no improvement or a drop in success rate when the reference is supplied.","supporting_citations":[],"review_version":1}