{"id":"9491da22-75d3-4dab-85f5-91d5452a1ee1","arxiv_id":"2502.03317","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A contact-aware trajectory optimizer, CAMP, plans paths where robots push movable objects, raising navigation and rearrangement success rates versus a contact-avoidance baseline.","lead":"This paper proposes a motion planning method that lets a mobile robot intentionally push movable objects, treating contact as an optimization constraint rather than something to avoid at all costs. In simulations and real-world trials, the method beats a contact-avoidance baseline on navigation and rearrangement tasks in crowded environments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The contact model in Appendix A drops ground friction, so planned object trajectories supporting the feasibility claim are physically inconsistent; Table II object-position errors (means 0.13-0.20 m, max 0.51 m) are the expected symptom.","rationale":"The reader identified the same assumption, and I agree it is the weakest load-bearing point. The paper's strongest claim has two parts: improved success in NAMO/RAMO and feasible trajectories. The success-rate part is supported by NAMO simulation and a subset of RAMO (though Table I shows no baseline for RAMO). The feasibility part is supported only by the real-world tracking data and the physics model in Appendix A. That model is not an unnecessary simplification; it changes whether the optimization variables are physically consistent. A trajectory with zero contact force while the robot and object move together at constant velocity is exactly the kind of solution the complementarity formulation can produce once ground friction is omitted. The real-world object tracking errors, roughly an order of magnitude larger than the robot tracking errors, are direct evidence that the executed object motion does not match the planned object motion. This does not mean the method is useless: the robot still completed tasks, and the NAMO success-rate comparison is plausible. But the feasibility claim as stated is not yet established, which is precisely the reader's conditional verdict. Therefore the verdict remains conditional; no adjustment is needed. Secondary issues, such as the missing RAMO baseline comparison and the lack of released code, reinforce the need for the concrete test rather than changing the verdict.","tokens_in":14829,"tokens_out":7655,"duration_ms":81874,"concrete_test":"Re-implement or run the published CAMP trajectories in a rigid-body simulator with a Coulomb-friction ground plane and a pusher-slider contact model (e.g., MuJoCo or pybullet), using realistic object masses and floor friction. Compare the executed object trajectory and final pose with the planned object trajectory; if the deviation exceeds the task tolerance, or if the planned zero-force constant-velocity push is infeasible, the feasibility claim fails. A complementary analytical check: augment h_dyn with the limit-surface ground-friction term and recompute the complementarity residual and required contact force for the trajectories reported in Table I and the real experiments.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central feasibility claim depends on the contact dynamics in Sec. V-B4/V-B5 and Appendix A. The dynamics constraint h_dyn = M*C_xx_ddot - J^T lambda + f_D = 0 has no ground-friction term, and Appendix A explicitly says that under the quasi-static approximation the frictional force between objects and the ground can be ignored. In the pusher-slider setting cited by the paper, ground friction is the load that the contact force must overcome; ignoring it lets the optimizer declare a constant-velocity push feasible with zero contact force (relative velocity zero and force magnitude zero simultaneously satisfy h_comp). The planned object trajectory then need not correspond to any physically realizable push. Real-world data in Table II are consistent with this mismatch: robot position tracking error is about 0.015 m, but movable-object position error is 0.13-0.20 m on average and up to 0.51 m. The paper does not report object mass, floor material, or friction coefficient, so the regime in which the approximation holds is not established. This is a load-bearing soft spot because the claim that generated trajectories are feasible is specifically validated by these experiments and by the contact model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes CAMP, an optimization-based contact-aware motion planner for mobile robots operating among movable objects. Contact between the robot and movable objects is encoded as complementarity constraints (Section V-B5), and the resulting nonconvex trajectory optimization problem is solved with an augmented Lagrangian method (Section V-C), initialized by front-end path searches for NAMO and RAMO tasks (Section IV). The method is evaluated in randomized NAMO and RAMO simulations against the GCOPTER baseline (Table I) and in real-world experiments with cylinder and cube objects, with tracking errors reported in Table II. The paper claims that CAMP expands the robot's reachable space, substantially improves task success rates, and produces physically feasible contact trajectories.","tokens_in":15118,"tokens_out":7705,"duration_ms":79065,"significance":"If the modeling gaps identified below are addressed, CAMP would be a valuable step toward treating deliberate contact with movable objects as a first-class citizen in mobile robot trajectory optimization. The ALM formulation is a sensible choice for complementarity-constrained programs, the NAMO comparison is informative, and the real-world experiments provide a useful sanity check. The paper also states an intention to open-source the code, which would help reproducibility. However, the contact dynamics and complementarity model are under-specified, the ground-friction approximation in Appendix A undermines the physical feasibility claim as stated, and the RAMO baseline and success-rate definitions are missing. These issues are load-bearing for the central claims, so the paper needs a revision before it can be accepted.","major_comments":[{"comment":"The decision variables in Eq. (6) are listed as x = [p, o_1, ..., o_N, t_1, ..., t_M], but the dynamics constraint h_dyn in Section V-B4 contains lambda, the contact impulse force magnitudes, and Section V-B5 uses f_C in the contact constraints without giving f_C as an explicit function of C_x and t. The Jacobian J is only described as being determined during collision detection. As written, h_dyn and h_comp are not evaluable functions of the decision variables, so the optimizer cannot enforce them. Please add lambda (or an explicit contact-force law such as a penalty model or friction-cone inequality) to the problem formulation, or state precisely how lambda and f_C are computed from C_x. This is load-bearing because the paper's feasibility claim rests on these constraints.","section":"Sec. V-A / V-B4"},{"comment":"Appendix A states that, under the quasi-static assumption, \"we can ignore the frictional force between the objects and the ground.\" This is not the pusher-slider quasistatic approximation used in the cited literature: in [31] and [32], the limit surface is precisely the mapping from the contact force required to overcome ground friction to the resulting object velocity. If ground friction is omitted from h_dyn, then a constant-velocity plan with zero contact force satisfies h_dyn (with air drag neglected) and h_comp (zero force), so the planned object motion is not anchored to a physically realizable push. The movable-object tracking errors in Table II (means 0.13-0.20 m, maxima up to 0.51 m) are consistent with this model mismatch. Please either include a ground-friction term (e.g., a Coulomb friction cone or limit surface) in h_dyn, or restrict and validate the approximation by reporting object mass, floor material, and friction coefficients and by showing that the tracking errors remain within a task-specific tolerance.","section":"Appendix A / Sec. V-B4"},{"comment":"Table I reports no GCOPTER baseline for the RAMO tasks (the RAMO rows show \"----\"), but the abstract claims that CAMP yields \"a significant improvement in the success rate of two types of fundamental tasks.\" The success-rate improvement is therefore supported only for NAMO. In addition, \"success rate\" is never defined: the reader cannot tell whether it means planner convergence, constraint satisfaction at sampled points, or physical task completion. Please define the metric and provide the missing RAMO baseline comparison, or restrict the claim to NAMO. Please also report what constitutes a failed trial (e.g., timeout, constraint violation, or execution failure) and where failures occur.","section":"Table I / Abstract"},{"comment":"The inequality constraints g_i,j = -||dot-O_i - dot-O_j|| <= 0 and g_k = -||f_C(\\ddot O_k)|| <= 0 are automatically satisfied by every trajectory because norms are nonnegative; the actual contact condition is carried entirely by the equality h_comp. However, the equality uses only magnitudes, so it cannot distinguish normal velocity from tangential sliding velocity or normal force from friction force. This loses the directional information needed for sliding contact, which is central to pushing an object along the floor. Please replace the norm-based complementarity by a component-wise formulation (normal gap/velocity versus normal force, and tangential velocity versus friction force) or state and justify the modeling assumptions under which the magnitude version remains sufficient for the NAMO and RAMO scenarios.","section":"Sec. V-B5"}],"minor_comments":[{"comment":"The notation s = n*(N+1) does not match the displayed decision vector x = [p, o_1, ..., o_N, t_1, ..., t_M]; please clarify the intended dimensions.","section":"Sec. V-A"},{"comment":"The RAMO rows leave the GCOPTER entries as \"----\" without explanation; please add a caption note or state clearly in the text that no RAMO baseline was run.","section":"Table I"},{"comment":"The paper says the code will be open-sourced, but no repository URL or availability statement is provided; please include one if the code is available at the time of the revised submission.","section":"Abstract / Contributions"},{"comment":"The reported success rate increases as the number of movable objects grows (Table I, NAMO rows, 75% to 100%), which is counterintuitive and not explained; a short discussion of why more movable objects make the task easier for CAMP would help the reader interpret the comparison.","section":"Sec. VI-A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript draws heavily on the authors' prior work (GCOPTER/MINCO [1]) for both the trajectory basis and the baseline, which is understandable but makes the missing RAMO baseline all the more noticeable. The central technical gap is the contact model, not the ALM optimization machinery; I believe the issues can be fixed with a careful reformulation and additional validation, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Xi, you should know: this paper does something real, but its central feasibility claim rests on a contact model that is too loosely specified and physically questionable. It is a solid integration, not a breakthrough, and it deserves a serious referee who will push for revisions.\n\nWhat is actually new: CAMP treats movable objects as free agents in the front-end and folds contact into a single MINCO-based trajectory optimization via complementarity constraints solved with ALM. That framing is genuinely useful. The NAMO simulation wins over GCOPTER are consistent and not trivial, and the real-world experiments do show the planner can produce push trajectories that execute in a lab setting. The trajectory customization section is a nice demonstration of flexibility.\n\nNow the soft spots, and the stress-test note is on target. Appendix A drops ground friction, citing quasi-static dominance of robot-object friction. But in pusher-slider, ground friction is exactly what the contact force must overcome. With h_dyn missing a ground-friction term, a constant-velocity push with zero contact force can satisfy the complementarity condition, so the optimizer may declare physically meaningless trajectories feasible. The Table II object errors, mean 0.13-0.20 m and max 0.51 m against a robot tracking error of 0.015 m, are consistent with that model mismatch. Also, lambda never appears in the decision variable list, leaving the dynamics constraint under-specified. The inequality forms of the complementarity constraints are vacuous norms; the product equality does the work, which is a minor presentation issue and not a fatal flaw. Success rate is never defined, and despite the abstract's promise, no code link is provided. These are all fixable, but they need fixing before the feasibility claim is credible.\n\nThe paper is worth a serious referee. It is a legitimate step for NAMO/RAMO planning, the sim results are plausible, and the approach can be built upon. Send it to review, expect major revisions, and ask for the code and a clearer contact model.","headline":"A useful integration of complementarity constraints into MINCO-based planning, but the friction model is too hand-wavy to support the feasibility claim.","tokens_in":15614,"tokens_out":2583,"would_cite":true,"duration_ms":24473,"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":"CAMP, a contact-aware motion planner, lets a mobile robot push movable objects during planned trajectories, expanding its reachable space and raising task success rates in simulation and real-world tests.","keywords":["contact-aware motion planning","complementarity constraints","augmented Lagrangian method","navigation among movable objects","rearrangement of movable objects","trajectory optimization","mobile robot","non-prehensile manipulation"],"falsifier":"Run the same planned pushing trajectory on floors with different friction, such as smooth tile versus carpet, or with objects of different mass, and measure the pushed object's path; if the deviation from the planned object path scales with ground friction or mass and exceeds the mean 0.13-0.20 meter tracking errors reported in Table II, the no-ground-friction quasi-static premise is falsified.","tokens_in":14652,"feed_emoji":"🤖","tokens_out":7801,"duration_ms":67600,"temperature":0.7,"pith_summary":"The paper argues that mobile robot motion planning should not treat every object as a static obstacle. It introduces CAMP, a trajectory optimization framework in which contact between the robot and movable objects is planned rather than avoided. Contact is encoded as complementarity constraints, and the resulting optimization is solved with an augmented Lagrangian method. In simulations, CAMP raises average success in navigation among movable objects (NAMO) to 95% from 54.3% for a collision-free baseline, and achieves 95-100% success on rearrangement of movable objects (RAMO). Real-world tests show the planned pushing trajectories are executable, which matters because deliberate contact lets robots move through cluttered human environments.","feed_headline":"Robots that push clutter instead of avoiding it complete far more tasks","feed_subtitle":"New planner turns movable objects into pushable aids, lifting success from 54% to 95% in cluttered navigation tests.","key_machinery":"The load-bearing object is the complementarity constraint on contact, written as $\\lVert \\dot{O}_i(t)-\\dot{O}_j(t)\\rVert_2 \\cdot \\lVert f_C(\\ddot{O}_k(t))\\rVert_2 = 0$ with both factors constrained non-negative: agents may move apart or transmit contact force, but not both at the same point. This expresses non-penetration and stick/slip behavior without prescribing a mode schedule in advance. The trajectory parameterization is piecewise polynomial with minimum control effort, built from the state sequence q and time allocation T; collision avoidance uses GJK distance computation and an ESDF field; the dynamics for movable objects are $M\\ddot{C}_x - J^T\\lambda + f_D = 0$; and the augmented Lagrangian method, using L-BFGS with a Lewis-Overton line search, solves the resulting optimization problem with complementarity constraints.","core_discovery":"The paper tries to establish that a robot's useful workspace expands when the planner is permitted to make contact with movable objects, and that such contact can be made predictable enough to plan as part of an optimization. It encodes contact as complementarity: for any robot-object pair, the relative speed and the contact force cannot both be nonzero, so the agents either separate or push while the contact sticks. This constraint is embedded in an ALM-based trajectory optimization over polynomial trajectories, whose decision variables are waypoints plus segment times. The reported results are that NAMO success rises from a 54.3% baseline average to 95%, RAMO tasks succeed 95-100% with back-end optimization times on the order of seconds, and experiments with a real omnidirectional robot pushing cylinders and cubes produce feasible trajectories. The authors claim this demonstrates a general contact-aware planning paradigm rather than a task-specific controller.","pith_inferences":["Inference: the paper's contribution is mostly at the trajectory level; task-level contact decisions still come from a front-end search on a masked map, so coupling CAMP with learned or semantic object-choice policies is a natural next step.","Inference: because the contact model ignores ground friction under a quasi-static assumption, the method should degrade on high-friction floors or with heavy objects, and adding ground-friction estimation or closed-loop correction is a testable extension.","Inference: the complementarity-plus-ALM formulation is not restricted to mobile robots; it could transfer to manipulation arms or legged robots that also plan through intentional contact, provided the same convex-shape collision geometry is available.","Inference: the reported object tracking errors, with maxima near half a meter, suggest that the planned contact model is only approximately correct; fusing force or tactile sensing during execution would be a direct way to close that gap."],"forward_implications":["In scenes with movable objects, a robot can navigate through spaces that collision-free planners treat as blocked, because the planner can choose to push an object aside instead of routing around it.","NAMO success rate rises as the number of movable objects grows, with success exceeding 90% when more than two movable objects are present, suggesting the method converts clutter into usable degrees of freedom.","RAMO tasks with cylinder and cube objects can be planned with back-end times of roughly 3 to 4 seconds and 95-100% simulation success, enabling rearrangement rather than only avoidance.","The same framework can be customized by adding objective terms, such as a preference for the pushed object's final pose: setting that weight to one makes the robot leave the object displaced instead of pushing it back.","Long-distance, long-duration pushing trajectories are executable in the real world, with movable-object position tracking errors averaging 0.13 to 0.20 meters and maxima up to 0.51 meters."],"supporting_citations":[{"why":"Supplies the MINCO polynomial trajectory parameterization used for the robot and movable objects, and the GCOPTER algorithm used as the collision-free baseline in comparisons.","marker":"[1]"},{"why":"Defines navigation among movable objects and the earlier discrete action-searching approach that CAMP extends to continuous, optimization-based trajectories.","marker":"[5]"},{"why":"Introduces direct trajectory optimization through contact using complementarity constraints, the conceptual basis for CAMP's contact formulation.","marker":"[7]"},{"why":"Provides the augmented Lagrangian method framework and convergence theory that CAMP uses as its optimization backbone.","marker":"[11]"},{"why":"Supplies the GJK algorithm used to detect collisions between the robot and movable objects for the collision-avoidance constraints.","marker":"[22]"},{"why":"Provides the limit surface concept mapping friction force to object velocity, part of the quasi-static contact model used in the appendix.","marker":"[31]"},{"why":"Gives the pusher-slider quasi-static analysis with limit surface and motion cone that underlies the contact friction assumptions.","marker":"[32]"},{"why":"Supports the claim that augmented Lagrangian methods converge for optimization problems with degenerate and complementarity constraints, justifying the solver choice.","marker":"[36]"},{"why":"Supplies the contact and friction simulation formalism, including gap functions and complementarity, used to write the non-penetration dynamics and contact constraints.","marker":"[39]"}],"fun_headline_variants":["Pushing clutter beats avoiding it: robots expand workspace","Robots that push, not dodge, achieve 95% task success","Contact-aware planner lets robots push obstacles to finish tasks","Move obstacles instead of avoiding them: new robot planner","Pushing movable objects as aids: robot planner boosts reach"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The contact model assumes the friction between the robot and a pushed object is much stronger than the friction between the object and the ground, so the ground's grip can be ignored during low-speed pushing.","fun_headline_variants_meta":{"raw":{"variants":["Pushing clutter beats avoiding it: robots expand workspace","Robots that push, not dodge, achieve 95% task success","Contact-aware planner lets robots push obstacles to finish tasks","Move obstacles instead of avoiding them: new robot planner","Pushing movable objects as aids: robot planner boosts reach"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000555,"raw_usage":{"total_tokens":2615,"prompt_tokens":887,"completion_tokens":1728,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":1646}},"tokens_in":503,"tokens_out":1728,"duration_ms":11277,"temperature":1.0,"reasoning_tokens":1646,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T05:08:29.791397+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same planned pushing trajectory on floors with different friction, such as smooth tile versus carpet, or with objects of different mass, and measure the pushed object's path; if the deviation from the planned object path scales with ground friction or mass and exceeds the mean 0.13-0.20 meter tracking errors reported in Table II, the no-ground-friction quasi-static premise is falsified.","supporting_citations":[{"cited_title":"Geometrically constrained trajectory optimization for multicopters,","cited_arxiv_id":null,"evidence_quote":"Supplies the MINCO polynomial trajectory parameterization used for the robot and movable objects, and the GCOPTER algorithm used as the collision-free baseline in comparisons."},{"cited_title":"Navigation among mov- able obstacles: Real-time reasoning in complex envi- ronments,","cited_arxiv_id":null,"evidence_quote":"Defines navigation among movable objects and the earlier discrete action-searching approach that CAMP extends to continuous, optimization-based trajectories."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the augmented Lagrangian method framework and convergence theory that CAMP uses as its optimization backbone."},{"cited_title":"Planar slid- ing with dry friction part 1. limit surface and moment function,","cited_arxiv_id":null,"evidence_quote":"Provides the limit surface concept mapping friction force to object velocity, part of the quasi-static contact model used in the appendix."},{"cited_title":"Feedback control of the pusher-slider system: A story of hybrid and under- actuated contact dynamics,","cited_arxiv_id":null,"evidence_quote":"Gives the pusher-slider quasi-static analysis with limit surface and motion cone that underlies the contact friction assumptions."},{"cited_title":"Global convergence of augmented lagrangian methods applied to optimization problems with degenerate constraints, including problems with complementarity constraints,","cited_arxiv_id":null,"evidence_quote":"Supports the claim that augmented Lagrangian methods converge for optimization problems with degenerate and complementarity constraints, justifying the solver choice."},{"cited_title":"Contact and friction simulation for computer graphics,","cited_arxiv_id":null,"evidence_quote":"Supplies the contact and friction simulation formalism, including gap functions and complementarity, used to write the non-penetration dynamics and contact constraints."}],"review_version":1}