{"id":"d72962cf-3f9f-48a4-b9f8-0cf6f4b7398a","arxiv_id":"1908.05380","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A dynamic-aware 'residual force polytope' is introduced to represent the task-space forces a robot can resist during a planned motion, and it is used as a trajectory optimization objective to improve robustness.","lead":"A team from Edinburgh and Oxford proposes a way to compute, for any planned robot motion, the set of external forces the robot could still resist without exceeding its motor limits. The paper shows that optimizing trajectories for this set, called the residual force polytope, yields motions that appear more robust to pushes than standard objectives, though the main test uses the same measure being optimized.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim rests on a polytope that may not be the true admissible-force set: Eq. (2) maps the residual torque box through the pseudo-inverse of J^T, which over-estimates feasible forces for a 7-DoF arm (n>m); the same object is used as objective and evaluation metric.","rationale":"The reader correctly saw that the evaluation metric is the same as the objective and is not independently validated. My pass identifies a more specific mechanism: for the redundant KUKA arm the linear map from residual torques to end-effector forces is not invertible, and the paper's pseudo-inverse construction (Eq. 2, Section 4) maps the entire residual torque box into task space, including torque components in the null space of J^T. Those components cannot be balanced by an end-effector force while keeping the nominal motion unchanged, so the computed polytope can be strictly larger than the true admissible set. This directly contradicts the paper's 'exact' claim and means both gE and the Section 7.2 metric may reward configurations whose apparent force margin comes from null-space torque. The counterexample with J^T=[1,2]^T is enough to show the two objects differ in general; the recomputation on the published KUKA trajectories would show whether the reported ranking survives. I therefore keep the reader's CONDITIONAL verdict: the paper is promising and the polytope idea is useful, but acceptance should require the corrected H-rep computation and either a closed-loop simulation or a clear statement that the robustness claim is about the computed margin only.","tokens_in":15320,"tokens_out":14075,"duration_ms":149323,"concrete_test":"Recompute Figure 7 with the H-rep definition instead of the pseudo-inverse image: for each mesh point, construct Q_k = { f : τ_lb - τ_nom(q_k,v_k,˙v_k) ≤ J_e(q_k)^T f ≤ τ_ub - τ_nom(q_k,v_k,˙v_k) }, and compute the origin-centered inscribed-ball radius r_k = min_i b_i / ||a_i||_2 from the half-space inequalities; then replot the gA–gE curves. Sanity check: for a 2-DoF arm with J^T = [1,2]^T and unit torque limits, the pseudo-inverse image radius is 0.6 while this H-rep radius is 0.5. If the corrected KUKA curves change materially, or if gE no longer dominates, the central robustness claim is an artifact of the over-approximation.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The most load-bearing assumption is that the set computed in Sections 3–4 is exactly the set of task-space forces the robot can resist while tracking the nominal motion. For the KUKA experiments, n=7 and m<7, so J_e^T is rectangular and the paper inverts it with the Moore–Penrose pseudoinverse (Section 3.2, Eq. (2)) before transforming the residual torque polytope (Section 4). This is not equivalent to the correct admissible set Q={f: τ_lb ≤ τ_nom + J_e(q)^T f ≤ τ_ub}. A residual torque with a null-space component produces no end-effector force, yet its pseudo-inverse image is a nonzero force whose induced torque J_e^T (J_e^T)^+ τ is only the projection of the original torque and may leave the feasible box. Hence the residual force polytope can be strictly larger than the true set, contradicting the 'exact representation' claimed in the Introduction. Because the same over-approximation is used both in objective gE (Eq. 18) and in the evaluation metric of Section 7.2, Figure 7 may rank trajectories by a quantity that does not correspond to true disturbance-rejection capability. No closed-loop simulation, hardware experiment, or released code checks the polytope against actual behavior, so the central robustness claim is not independently supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces the residual force polytope, defined as the image under the (pseudo)inverse of the end-effector Jacobian transpose of the box of joint torques remaining after subtracting the nominal torques of a planned trajectory. It uses the Chebyshev-ball radius of this polytope as an isotropic robustness metric, a cone-intersection volume as a directional robustness metric, and embeds both as objective functions in a direct-transcription trajectory optimization for a 7-DoF KUKA arm. The experiments compare objectives gA-gF, report that gE (residual force polytope radius) yields trajectories with larger admissible force magnitudes, benchmark interior-point versus SQP methods, and test directional robustness of gE versus gF under a synthetic impulse.","tokens_in":15598,"tokens_out":10083,"duration_ms":105376,"significance":"If the representation were exact and the evaluation independent, this would be a useful offline planning and design tool: it replaces ellipsoidal manipulability approximations with a convex polytope that respects actuation limits, accounts for the torques required by the motion, and yields interpretable margin metrics. The paper is clearly written, the polytope/LP machinery is standard, and the computational benchmarks in Tables 1-4 are a useful practical contribution. However, the headline claim is currently supported only by evaluating trajectories with the same quantity that objective gE maximizes, and the polytope construction itself over-approximates the true admissible-force set for a redundant arm. These issues must be resolved before the robustness claim can be accepted.","major_comments":[{"comment":"For the redundant case n > m considered in the experiments, the residual force polytope P'_F = (J_e^T)^+ P'_tau is not equal to the true set Q = { f in R^m : tau_lb <= tau_nom + J_e(q)^T f <= tau_ub }. The pseudoinverse discards the component of each residual torque orthogonal to range(J_e^T), and a force is in the pseudo-inverse image exactly when the projection of the residual torque onto range(J_e^T) lies in the residual torque box. This condition is weaker than requiring J_e(q)^T f itself to lie in the box, because the projection of a box onto a subspace can be strictly larger than the intersection of the box with that subspace. Hence P'_F can contain forces whose induced joint torques violate the actuation limits, contradicting the 'exact representation' claimed in the Conclusions and the claim in Section 2 that the approach determines 'the exact margins remaining before torque saturation occurs.' This over-approximation affects both objective gE and the evaluation metric in Section 7.2.","section":"Section 3.2, Eq. (2); Section 4"},{"comment":"The evaluation metric used to rank trajectories is identical to the objective being optimized. Equation (18) defines gE as the sum over mesh points of Br(P'_k), the radius of the largest ball centered at the end-effector and inscribed in the residual force polytope, and Section 7.2 states that each trajectory is evaluated by computing the same residual force polytope and taking the same largest inscribed ball radius. Consequently, Figure 7 demonstrates that maximizing a function improves that same function, which is a tautological comparison unless Br(P') is independently shown to predict closed-loop robustness to unknown external disturbances. The impulse test in Section 7.3 does not provide that independent validation: it compares only gE with gF, and it computes the required extra torques using Eq. (2), the same pseudoinverse mapping that defines the polytope. The paper should add an evaluation against the exact admissible set, a closed-loop simulation, or a hardware experiment.","section":"Section 7.2 and Eq. (18)"},{"comment":"The paper motivates robustness by sensor noise, signal delay, and model mismatches, but the experimental validation contains no feedback controller, no friction or noise model, and no hardware data. The planned torque margins are evaluated open-loop, so the central claim that the resulting trajectories are 'more robust to unknown external disturbances' is not independently supported. At minimum, a closed-loop torque-control simulation with a disturbance observer or a hardware disturbance-rejection test should be reported; Section 8.1's discussion of scalability does not substitute for such validation.","section":"Sections 7.3 and 8.1"}],"minor_comments":[{"comment":"The notation tau_lim in Eq. (15) is ambiguous: Eq. (3) defines per-joint scalar limits, so the objective should state explicitly whether tau_lim is a vector of per-joint limits and how asymmetric bounds are handled.","section":"Eq. (15)"},{"comment":"The symbol J_e^{-T} suggests invertibility, but for n > m the object used is the Moore-Penrose pseudoinverse; please write (J_e^T)^+ consistently and state the rank assumption needed for the pseudoinverse formula.","section":"Section 3.2, Eq. (2)"},{"comment":"The LP for the largest inscribed ball should state explicitly that the polytope is expressed in the end-effector frame and that the ball center is the origin of that frame; the phrase 'centered at the end-effector' is correct but less precise.","section":"Section 5.1"},{"comment":"The caption says the plots show objectives gA-gE, while Figure 5 includes panels for gF; consider clarifying that gF is omitted because it is direction-dependent and treated separately in Section 7.3.","section":"Figure 7"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a revised preprint whose central claim needs additional validation. The two load-bearing issues are the pseudoinverse over-approximation for redundant arms and the circular use of the same metric as objective and evaluation. Both are fixable within the scope of the paper, so I recommend major revision rather than rejection. I saw no indication of author misconduct or duplicate publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know upfront. First, the idea is genuinely new and useful: subtract the nominal torque trajectory from the actuation limits and map the remainder to task space, giving a dynamic-aware robustness measure that connects trajectory optimization to disturbance rejection. Second, the way they build that mapping for a redundant arm is mathematically wrong, and because the same object is both the cost and the evaluation metric, the main empirical claim doesn't yet stand.\n\nWhat's good: the paper is clear and honest. The residual force polytope (Sec. 4) and the two uncertainty models (ball, cone) are well motivated. The experiments compare six objectives under two NLP solvers, with and without payload, and the computational-cost breakdown is refreshingly candid. The admitted limits on scalability are accurate, and the related work is fair.\n\nNow the soft spots. The stress-test you sent is right. For the KUKA arm, n=7 and the task space is 3D linear forces, so J^T is rectangular. The correct set of end-effector forces consistent with the residual torque box is Q = {f : τ_nom + J^T f ∈ [τ_lb, τ_ub]}, an H-polyhedron in R^m. The paper instead maps the residual torque box through (J^T)^+ and calls the image exact. For non-square J, that image is generally a strict superset of Q: a torque with a null-space component projects to a smaller torque, and the corresponding force can require a torque outside the box. So the 'exact representation' claim in the Introduction and Conclusions is false. This matters doubly because gE maximizes the Chebyshev-ball radius of this over-approximated set, and Section 7.2 evaluates all trajectories with the same radius on the same over-approximated set. Figure 7's ranking of gE first is therefore circular, or at least not probative for true robustness. The impulse test in 7.3 is meant to be independent, but it says it computes required torques 'with equation (2)', which maps torque to force; the correct map from an external force to torque is τ = J^T f. If they used the pseudo-inverse in the wrong direction, that test is also invalid.\n\nWhat still holds: the conceptual move—account for the nominal motion's torque demand when measuring remaining capability—is sound and can be repaired by defining the residual force polytope as the H-rep set Q, or by acknowledging the over-approximation and validating against the exact set. The comparison framework (gA–gF, payload, solver choice) is a solid template for revision.\n\nBottom line: the paper deserves serious peer review but not in its current form. I'd send it to referees expecting major revision, and the pseudoinverse issue must be resolved or openly scoped before publication. Don't cite it yet as a correct method. It would make a good reading-group discussion about polytope mappings and evaluation circularity.","headline":"A useful dynamic-aware force polytope idea, but the redundant-arm construction over-approximates the true admissible-force set and is used as both the objective and the evaluation metric, so the central robustness claim is not yet supported.","tokens_in":16141,"tokens_out":6398,"would_cite":false,"duration_ms":63823,"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":"Optimizing trajectories over the residual force polytope yields motions that withstand larger unknown disturbances than standard objectives.","keywords":["residual force polytope","force polytope","robustness","trajectory optimization","polytopes","task-space forces","manipulability","robotic arms"],"falsifier":"Run a $g_E$-optimized trajectory and a torque-minimized ($g_A$) trajectory on the same robot, or in a simulation with a tracking controller, friction, and measurement noise, and apply an external force impulse of increasing magnitude from random directions to the end-effector, recording the largest impulse each motion rejects before torque saturation or a set tracking-error bound is exceeded; if the $g_A$ trajectory rejects as much as the $g_E$ trajectory despite its smaller inscribed-ball radii, the residual force polytope does not predict real robustness. A minimal check: compare the metric's predicted admissible force at a single trajectory point against the largest push the closed-loop system actually tolerates at that point.","tokens_in":15111,"feed_emoji":"🤖","tokens_out":17762,"duration_ms":145790,"temperature":0.7,"pith_summary":"The paper introduces the residual force polytope, an exact convex representation of the task-space forces a robot can still exert or resist at any instant of a dynamic trajectory without abandoning its planned motion: the set of forces reachable from the joint torques left over after the nominal torques of the motion are subtracted from the actuation limits, mapped through the inverse-transpose Jacobian $J_e^{-\\top}$ to the end-effector frame. Its central claim is that an optimization objective built from this polytope—maximizing the radius of the largest ball, centered at the end-effector, that fits inside it—produces trajectories that are more robust to unknown external disturbances than objectives based on torque minimization, manipulability maximization, or the classical quasi-static force polytope. The authors demonstrate this on a 7-DoF arm, where the residual-polytope objective beat all alternatives in admissible disturbance force over the whole trajectory, and they report that the cost of polytope operations restricts the method to offline planning for now.","feed_headline":"Residual torque margins yield more robust robot trajectories","feed_subtitle":"Optimizing over leftover force margins beats torque, manipulability, and static objectives in disturbance tests.","key_machinery":"The residual force polytope $P'(q,\\tau)$: the convex polytope of task-space forces obtained by mapping the torque margin $\\tau_{\\text{lim}} - \\tau$, i.e., the nominal torques of the motion subtracted from the joint actuation limits, through the inverse-transpose Jacobian $J_e^{-\\top}$ into the end-effector frame. It is the exact, dynamics-aware replacement for the quasi-static force polytope: because the nominal torques of each trajectory point are decision variables in the direct transcription, an objective built from $P'$ cannot merely inflate geometric manipulability but must preserve torque authority against the very motion being demanded. Two metric computations carry the evaluation. The first is the inscribed-ball radius, solved as the linear program that maximizes $r$ subject to $B(0,r) \\subseteq P'$ with the ball center fixed at the end-effector origin, which measures isotropic robustness to forces from unknown directions. The second is the volume of the intersection $P' \\cap C$ with an uncertainty cone $C$, which measures robustness to forces from a roughly known direction. The double-description method for converting between half-space and vertex representations drives the computational cost, which the paper measures at roughly 22–24 $\\mu$s per polytope construction but 6.3–7.4 ms per ball, intersection, or volume operation, and which restricts the objectives to offline planning.","core_discovery":"The paper's central object is the residual force polytope $P'(q,\\tau)$: given the equations of motion $M(q)\\dot v + h(q,v) = \\tau + J_e^\\top(q) f_{\\text{tip}}$, the residual torques $\\tau_{\\text{lim}} - \\tau_{\\text{nominal}}$ are mapped through $f_{\\text{tip}} = J_e^{-\\top}\\tau_{f_{\\text{tip}}}$ to obtain the convex set of end-effector forces that can be resisted without saturating any joint and without interfering with the nominal motion. From this object the paper extracts two scalar robustness metrics: the radius of the largest ball centered at the end-effector inscribed in the polytope—a linear program, i.e., a Chebyshev-center problem with the ball center fixed—for disturbances of unknown direction, and the volume of the polytope's intersection with a cone aligned to an expected force direction, for partially known disturbances. Both metrics are summed over all mesh points as objectives $g_E$ and $g_F$ in a direct-transcription optimal control problem for a 7-DoF KUKA LWR arm, alongside comparison objectives $g_A$–$g_D$ (torque minimization, residual-torque maximization, manipulability maximization, and the quasi-static force-polytope inscribed ball). The paper's central result, reported for both interior-point and active-set solvers, is that $g_E$ yields trajectories with greater admissible force magnitude over the entire trajectory than every alternative, improving on $g_D$ by $53.2\\pm6.52$ N without payload and $40.55\\pm21.37$ N with a 2 kg payload, while the cone-based $g_F$ gives still larger margins when the disturbance direction is correctly anticipated but saturates the actuators when the disturbance arrives from the opposite direction.","pith_inferences":["The paper validates its metric only on open-loop trajectories, so the natural unrun experiment is a closed-loop one: whether the inscribed-ball radius predicts realized disturbance rejection under a feedback controller with friction, noise, and tracking error is an open question, and the central claim's practical value hinges on it.","The cone model implies a continuum of specialization: any estimated distribution of disturbance directions—from contact geometry, terrain statistics, or prior trials—could be encoded in the cone's axis and aperture, letting a planner interpolate between the isotropic objective $g_E$ and the fully specialized $g_F$.","Because $P'(q,\\tau)$ encodes actuation limits, dynamics, and the task in one covariant set, the same metric could serve system co-design (e.g., actuator sizing or mounting-choice selection) by ranking hardware according to the worst-case force it can resist while executing a nominal motion—an application the paper names but does not demonstrate.","A surrogate model that approximates the residual polytope geometry would settle the scalability question: if an approximate polytope preserves the robustness ranking of $g_E$ over $g_A$–$g_D$ while cutting evaluation time from milliseconds to microseconds, the approach could move from offline to near-online planning."],"forward_implications":["Trajectories optimized with $g_E$, the sum of inscribed-ball radii in residual force polytopes, have greater admissible force against disturbances from any direction over the whole trajectory than trajectories from $g_A$–$g_D$, under both interior-point and active-set solvers, so robustness can be put into the plan rather than left to the controller.","Using the residual force polytope instead of the quasi-static force polytope in the same objective adds $53.2\\pm6.52$ N of admissible force without payload and $40.55\\pm21.37$ N with a 2 kg payload for the tested 1 s trajectories, showing that the static polytope overestimates the capability actually available while a motion is running.","Direction-specialized plans ($g_F$) beat isotropic plans ($g_E$) in margin when the disturbance comes as predicted, but saturate joint torques when it comes from the opposite direction, so isotropic planning is the safer default unless the disturbance direction is nearly certain.","The residual force polytope can be computed once per mesh point for any given trajectory, which makes it an offline evaluation tool for ranking arbitrary motions by robustness even when no optimization is involved.","The extra cost of the residual variant over the static one is small; the expensive operations are the polytope geometry routines themselves, so the paper's sketched remedies—evaluating the polytope less often or approximating its geometry—are the concrete routes toward online use."],"supporting_citations":[{"why":"Introduces the actuation wrench and feasible wrench polytopes for quasi-static legged robots, the precursor that the residual force polytope extends to dynamic scenarios and the template for objective g_D.","marker":"[4]"},{"why":"Defines the manipulability ellipsoid and its measure, which supply comparison objective g_C and the kinematic-only view of capability.","marker":"[7]"},{"why":"Provides the double-description method used to convert between half-space and vertex polytope representations, the dominant computational cost of the proposed objectives.","marker":"[15]"},{"why":"Defines the force polytope and scaled manipulability for redundant manipulators, the basis of the quasi-static comparison objective g_D.","marker":"[16]"},{"why":"Supplies the Chebyshev-center linear-programming formulation from which the largest inscribed-ball robustness metric is derived.","marker":"[17]"},{"why":"Provides the direct-transcription framework, including defect constraints, on which the trajectory optimization problem is built.","marker":"[18]"},{"why":"Supplies the interior-point and active-set solvers whose runs produce the reported comparisons across objectives and the convergence-time tables.","marker":"[25]"}],"fun_headline_variants":["Residual force polytope metric boosts robot trajectory disturbance resistance","Optimizing force polytope margins outperforms torque and manipulability goals","New force polytope objective yields more disturbance-robust robot paths","Residual force polytope: better trajectory robustness via leftover torque margins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the radius of the largest ball of end-effector forces that fits inside the residual force polytope is what real robustness to unknown disturbances consists of; the experiments test that radius only on the planned open-loop trajectories, with no feedback controller, hardware, friction, or sensor noise in the loop, so if that geometric radius does not predict how the closed-loop robot actually rejects disturbances, the comparison that supports the central claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["Residual force polytope metric boosts robot trajectory disturbance resistance","Optimizing force polytope margins outperforms torque and manipulability goals","New force polytope objective yields more disturbance-robust robot paths","Residual force polytope: better trajectory robustness via leftover torque margins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000313,"raw_usage":{"total_tokens":1848,"prompt_tokens":1083,"completion_tokens":765,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":699,"completion_tokens_details":{"reasoning_tokens":688}},"tokens_in":699,"tokens_out":765,"duration_ms":8748,"temperature":1.0,"reasoning_tokens":688,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:15:42.432557+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a $g_E$-optimized trajectory and a torque-minimized ($g_A$) trajectory on the same robot, or in a simulation with a tracking controller, friction, and measurement noise, and apply an external force impulse of increasing magnitude from random directions to the end-effector, recording the largest impulse each motion rejects before torque saturation or a set tracking-error bound is exceeded; if the $g_A$ trajectory rejects as much as the $g_E$ trajectory despite its smaller inscribed-ball radii, the residual force polytope does not predict real robustness. A minimal check: compare the metric's predicted admissible force at a single trajectory point against the largest push the closed-loop system actually tolerates at that point.","supporting_citations":[{"cited_title":"Orsolino, M","cited_arxiv_id":null,"evidence_quote":"Introduces the actuation wrench and feasible wrench polytopes for quasi-static legged robots, the precursor that the residual force polytope extends to dynamic scenarios and the template for objective g_D."},{"cited_title":"Yoshikawa, Manipulability of Robotic Mechanisms, The International Journal of Robotics Research (IJRR) 4 (2) (1985)","cited_arxiv_id":null,"evidence_quote":"Defines the manipulability ellipsoid and its measure, which supply comparison objective g_C and the kinematic-only view of capability."},{"cited_title":"Fukuda, A","cited_arxiv_id":null,"evidence_quote":"Provides the double-description method used to convert between half-space and vertex polytope representations, the dominant computational cost of the proposed objectives."},{"cited_title":"Chiacchio, Y","cited_arxiv_id":null,"evidence_quote":"Defines the force polytope and scaled manipulability for redundant manipulators, the basis of the quasi-static comparison objective g_D."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the direct-transcription framework, including defect constraints, on which the trajectory optimization problem is built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the interior-point and active-set solvers whose runs produce the reported comparisons across objectives and the convergence-time tables."}],"review_version":1}