{"id":"954544ae-92e1-43c3-a555-4b9a0abf682a","arxiv_id":"2608.00206","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A nonlinear MPC with a soft reachability penalty generates push/pull door-opening and traversal trajectories for a mobile manipulator, demonstrated in Isaac Sim and on hardware.","lead":"A motion planner for mobile manipulators folds driving and door-opening into one nonlinear optimization, handling both push and pull doors without modeling the arm. It relies on a soft penalty for arm reachability, and the authors demonstrate it in simulation and one hardware run.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Soft reachability penalty does not enforce the δdoor switch: the planner can command door rotation while the handle is outside the arm's workspace, so the 'ensured feasible' claim relies on an unstated finite weight.","rationale":"The paper proposes a low-dimensional single-NMPC planner with a soft reachability penalty and demonstrates push/pull traversal in Isaac Sim and one hardware experiment. The qualitative demos are real evidence that the idea can work in favorable cases, but the strongest claim—that the same continuous formulation is enough to generate executable trajectories because reachability is enforced—requires the soft penalty to reproduce the δdoor switch. The reader's weakest_assumption identifies exactly this point, and the full text supports it: the switched dynamics in (1) are replaced by a plain integrator in (14), and only the finite-weight penalty (6)/(9)/(12) remains. No hard constraint or weight value is given, so the central feasibility guarantee is not established. The collision-avoidance constraint (15) has additional issues (it ignores the arm and relies on a line-segment door model), but those are secondary to the reachability gap because door rotation while unreachable defeats the manipulation task outright. I therefore agree with the reader's CONDITIONAL verdict: the claims should be softened to 'demonstrated in the tested scenarios' or the authors should add a hard reachability constraint and report the penalty weights/quantitative violation metrics. The concrete test above would settle whether the penalty actually enforces the switch; until then, acceptance should remain conditional.","tokens_in":7905,"tokens_out":6875,"duration_ms":70073,"concrete_test":"Re-solve the same push-door scenario (Fig. 3) with the switching dynamics restored: replace the fourth-state equation in (14) by θdoor,k+1 = θdoor,k + δdoor,k ωdoor,k dt, with δdoor,k defined by Eq. (5) (1 iff r_min ≤ ||p_base,k − p_handle,k|| ≤ r_max). Compare the optimal θdoor profile and d_k time series with the soft-penalty NLP. Then sweep w_m (and w_manip) from, say, 10^0 to 10^6 in the soft-penalty problem. If any soft solution has |Δθdoor| > 0 at a step where d_k is outside [r_min, r_max], or if the violation does not go to zero as w_m → ∞, the penalty is not equivalent to the claimed switch and the 'ensured feasible' wording must be weakened. A minimal two-step instance with the base initially beyond r_max and a goal requiring door motion is enough to expose the trade-off.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that one continuous NLP jointly plans base motion and door opening so the planned door trajectory is executable by the arm. Executability requires the door to move only while the handle lies in the arm's reachable annulus [r_min, r_max]. The switched model (1)-(2) encodes this, but the final MPC (9)-(17) drops δdoor: the fourth-state update in (14) is a plain single integrator, θdoor_dot = ωdoor, and reachability remains only in the penalty w_m C_m = w_m w_manip (max{0,d^2−r_max^2}+max{0,r_min^2−d^2}) |ωdoor| (Eqs. 6, 9, 12). Nothing hard prevents the optimizer from rotating the door when d_k is outside the annulus; it merely pays a cost proportional to the violation times |ωdoor|. For finite w_m this can be outweighed by progress toward the goal, particularly if a small early door rotation reduces later terminal cost. The paper gives no value for w_m/w_manip and no bound showing the penalty dominates the objective. The sentence in Section 3 that trajectories are 'ensured to be feasible and collision-free' is therefore unsupported. The nonsmooth max terms also complicate the IPOPT solution. If this gap is realized, the arm IK cannot track the planned handle trajectory and the entire manipulation step fails.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a motion-planning framework for a mobile manipulator opening and traversing push and pull doors. The planner is formulated as a nonlinear MPC whose state includes the mobile-base pose and the door angle, with controls for base velocity and door angular velocity. The switched reachability condition from the motivating model is removed and replaced by a weighted soft penalty based on the distance between the base and the door handle. Collision avoidance with the door panel is imposed through a closest-point nonlinear constraint. The resulting NLP is solved once with IPOPT/CasADi, and the planned trajectory is tracked by a separate base MPC and differential IK. The paper reports Isaac Sim simulations and a single hardware experiment.","tokens_in":8285,"tokens_out":4002,"duration_ms":49414,"significance":"If the central claim were fully established, the contribution would be a modest but practically useful planning formulation: a single continuous optimization that handles push and pull doors without integer variables or full arm kinematics. The paper has some strengths: the geometric model is clearly explained, the use of a soft reachability penalty is an interesting way to avoid hybrid dynamics, and the demonstration includes both simulation and hardware. However, the load-bearing claim that generated trajectories are guaranteed feasible is not supported by the presented formulation, and the evaluation is anecdotal. The contribution is therefore plausible but not yet verified to the standard implied by the paper's wording.","major_comments":[{"comment":"The statement that “Trajectories generated by the nonlinear program are ensured to be feasible and collision-free” is not supported. In Eq. (14), the door dynamics are the plain single integrator θdoor_dot = ωdoor; the reachability condition δdoor from Eq. (1) is removed and appears only as the soft penalty w_m * C_manip in Eq. (9), with C_manip given by Eq. (6). Since C_manip is multiplied by |ωdoor|, the optimizer can rotate the door when d_k is outside [r_min, r_max] whenever the resulting progress toward the goal outweighs the finite penalty. No value or bound for w_m/w_manip is given, and no hard constraint prevents door motion outside the reachable annulus. Consequently, the downstream IK may be commanded to track a handle trajectory that the arm cannot execute. Please either restore a hard switch (e.g., complementarity or a bilinear constraint), provide a rigorous bound under whic","section":"Section 2.2, Eqs. (6), (9), (12), (14); Section 3, first sentence"},{"comment":"The abstract and Section 1.1 describe the robot and door as a “coupled dynamical system” and a “unified nonlinear model.” However, in Eq. (14) the base obeys unicycle dynamics and the door obeys a single integrator that is dynamically independent; the coupling is only through the cost term C_manip and the geometric constraint (15). This is a meaningful distinction: the contribution is a joint optimization over two dynamically decoupled subsystems with coupling in the objective/constraints, not a coupled dynamical model. Please revise the claims to match the actual formulation, or introduce true dynamic coupling if that is intended.","section":"Abstract, Section 1.1, Eq. (14)"},{"comment":"The experimental and simulation evaluation is not quantitative enough to support the claims of feasibility and versatility. No values are reported for r_min, r_max, w_m, w_manip, dt, N, Qf, R, or d_s, so the results are not reproducible from the paper. The hardware experiment appears to be a single successful run, and no tracking errors, success criteria, or statistics are given. For the central reachability concern, the paper should report the time history of d_k during the manipulation phase and show that door rotation occurs only when d_k ∈ [r_min, r_max]. Without such evidence, the planner's practical reliability and the claimed advantage over the switched formulation remain unsubstantiated.","section":"Section 3 and Figures 3–7"}],"minor_comments":[{"comment":"Eq. (12) defines Cm(xk) = w_manip * C_manip(xk, u_k), while Eq. (9) already multiplies Cm by w_m. This double weighting is confusing; either remove one weight or explain the role of each. Also, C_manip in Eq. (6) is written without a subscript k and without u_k as an argument, though it clearly depends on d_k and |ωdoor|.","section":"Eq. (12)"},{"comment":"The penalty contains nonsmooth max terms. IPOPT assumes smoothness, and the nonsmoothness may cause convergence issues or suboptimal local solutions. Consider reformulating with slack variables or a smooth approximation, and report solver status.","section":"Eq. (6)"},{"comment":"Calling the planner “MPC” is somewhat misleading because the optimization is solved once in open loop (“only needs to be solved once”) rather than in a receding horizon. The tracking controller in Eq. (19) is the receding-horizon part. Clarify terminology.","section":"Section 2.2, Eq. (14)"},{"comment":"The link to videos is helpful but the paper itself should include quantitative data: trajectory duration, computation time, base tracking error, and final door angle. The current statement “takes a few seconds” is too vague.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a control-applications venue, and I see no circularity or authorship concern. The main issue is that the central claim of guaranteed feasibility rests on a soft penalty with unreported weights and no verification that the reachability condition held during execution. This is fixable by tightening the formulation or by supplying the missing evidence, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a solid, incremental paper on MPC-based door traversal for mobile manipulators. The genuinely new bit is replacing the integer switch that historically gates door rotation (δdoor in the switched model) with a continuous soft penalty that only discourages door motion when the handle is outside the arm's reachable annulus. That keeps the whole plan a single continuous NLP, and the same formulation handles push and pull doors. I believe the authors, and the short hardware demo on a Realman is genuine evidence that the idea works in at least one scenario.\n\nThe main weakness is exactly the stress-test concern. The penalty in Eq. (6) is soft; nothing hard prevents the optimizer from rotating the door when the base is out of reach, if that reduces the terminal cost enough. No value for w_m or w_manip is given, and no bound shows the penalty dominates the objective. So the sentence in Section 3 saying the trajectories are 'ensured to be feasible and collision-free' is not supported. The collision constraint also only checks the base footprint, not the arm, and the 'coupled' dynamics are really independent — coupling is only through cost and constraints. These are fixable, but the claims need to be softened or the formulation needs a hard version of the reachability constraint.\n\nWhat's missing is any quantitative evaluation. There are no tables, no baselines, no success rates, and no timing. The evidence is a few snapshots and one hardware run. That's fine for a proof-of-concept, but it undercuts the stronger language in the abstract.\n\nFor a reader working on mobile manipulation planning, this is a useful example of how to shave a hybrid switch into a soft penalty. I'd send it to reviewers — it deserves a real referee — but I'd expect the authors to either relax the 'ensured' language or add hard constraints and numbers.\n\nBest","headline":"A plausible but overclaimed single-NMPC planner for door traversal; the 'ensured feasible' line rests on a soft penalty with no weight or bound given.","tokens_in":8743,"tokens_out":4425,"would_cite":false,"duration_ms":46115,"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":"Mobile manipulators can be planned to open and traverse doors as a single continuous MPC problem, with a soft reachability penalty standing in for arm kinematics.","keywords":["motion planning","model predictive control","mobile manipulator","door traversal","nonlinear optimization","manipulation feasibility","collision avoidance","push-pull doors"],"falsifier":"Examine the optimizer's output: if any planned time step has nonzero door angular velocity while the base-handle distance is outside [r_min, r_max], the manipulation-feasibility claim is violated. Concretely, re-solve the problem with the reachability weight reduced by an order of magnitude and observe whether the door swings while the base remains far from the handle.","tokens_in":7826,"feed_emoji":"🚪","tokens_out":3902,"duration_ms":44533,"temperature":0.7,"pith_summary":"This paper claims that opening and traversing a door can be planned as a single continuous nonlinear optimization rather than a staged sequence of subtasks. The robot base and the door are treated as one dynamical system, with the door angle as a state and door angular velocity as a control. Arm reachability is encoded not by modeling the arm, but by a penalty that allows door rotation only when the base is within an annular region around the handle. The planner is validated in simulation and on a physical mobile manipulator for both push and pull doors. If this holds, it means door crossing becomes a routine input to an MPC solver instead of a hand-crafted skill.","feed_headline":"One MPC planner opens both push and pull doors in a single optimization","feed_subtitle":"A reachability penalty replaces integer switches, so base motion and door angle are planned together for both door types.","key_machinery":"The load-bearing objects are the coupled state vector x=[x,y,θ_base,θ_door] with inputs [v,ω_base,ω_door]; the manipulation penalty C_manip = (max{0,d²-r²_max}+max{0,r²_min-d²})·|ω_door|; and the geometric door-clearance constraint built from the scalar projection t_k = (w_k·v_k)/||v_k||² and the closest-point function p_closest,k. Together they remove the need for integer variables and for explicit arm kinematic constraints in the planner.","core_discovery":"The central claim is that the switched dynamics with an integer reachability flag can be replaced by a soft penalty without losing feasibility. Inside the nonlinear program, the door obeys a simple single-integrator dynamic at every step, and the optimizer is steered toward opening the door only when the handle distance lies between an inner and outer reachability radius. A projection-based constraint keeps the base outside the door panel plus a safety margin. The same formulation is shown to produce a pushing trajectory and a pulling trajectory, with the arm's job reduced to tracking the planned door angle through differential inverse kinematics.","pith_inferences":["Because reachability is only a soft penalty, the planner has no hard guarantee against commanding door motion out of reach; a hard constraint or a small mixed-integer program would close that gap at extra solve cost.","The same coupling idea might apply to other articulated objects (cabinets, valves, drawers) by replacing the door-angle state with the articulation angle and reusing the annular reachability penalty.","A natural stress test would be a door with stiction or an unexpected obstruction: the geometric model here has no force or torque model, so the planner may need to be embedded in a feedback loop rather than executed open-loop.","Retargeting to arms with non-annular workspaces (e.g., long thin arms) may degrade because the penalty assumes radial symmetry; one could substitute a measured reachability map."],"forward_implications":["Door traversal can be planned as one nonlinear program solved once in a few seconds, replacing staged state machines or behavior trees.","The same planner covers push doors and pull doors without switching formulations, because the door direction emerges from the goal state and costs.","Arm-specific detail reduces to two radii (r_min, r_max); the planner can be retargeted to a new manipulator by changing these numbers.","Additional obstacles can be added as projection-based distance constraints, so the method should extend beyond the empty-corridor setting.","The planned door-angle trajectory provides a direct reference for a Jacobian-based arm tracker, bridging planning and low-level execution."],"fun_headline_variants":["Soft penalty replaces integer switch for door traversal MPC","One MPC unifies push and pull door planning via penalty","MPC planner handles both door types without explicit switching","Reachability penalty enables single MPC for door negotiation","Soft reachability constraint unifies push/pull door MPC"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim that the planned door motion is executable rests on the soft penalty being weighted heavily enough that the optimizer never chooses to rotate the door while the base is outside the reachable annulus; no hard constraint enforces this, and no value or guarantee is supplied.","fun_headline_variants_meta":{"raw":{"variants":["Soft penalty replaces integer switch for door traversal MPC","One MPC unifies push and pull door planning via penalty","MPC planner handles both door types without explicit switching","Reachability penalty enables single MPC for door negotiation","Soft reachability constraint unifies push/pull door MPC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000447,"raw_usage":{"total_tokens":2020,"prompt_tokens":598,"completion_tokens":1422,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":342,"completion_tokens_details":{"reasoning_tokens":1345}},"tokens_in":342,"tokens_out":1422,"duration_ms":10456,"temperature":1.0,"reasoning_tokens":1345,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T00:59:51.354520+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Examine the optimizer's output: if any planned time step has nonzero door angular velocity while the base-handle distance is outside [r_min, r_max], the manipulation-feasibility claim is violated. Concretely, re-solve the problem with the reachability weight reduced by an order of magnitude and observe whether the door swings while the base remains far from the handle.","supporting_citations":[],"review_version":1}