{"id":"4e3f916a-5733-4340-ae28-ca37948c9af1","arxiv_id":"2504.18667","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A planning and control stack that transports passive objects in microgravity through robot impacts, using STL-based MILP planning, online replanning, and MPC, with hardware validation.","lead":"This paper shows a team of free-flying robots can move a floating object in space by bumping into it, rather than grabbing it. The authors provide a full planning and control stack, tested in a high-fidelity simulator and on a physical freeflyer platform, with code and videos online.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Hardware validation is confounded: the 'passive' object runs a velocity-keeping MPC, so impacts are not the only control of object motion in the experiments supporting the central claim.","rationale":"The reader's conditional verdict is appropriate, but I reach it through a different, more direct concern. The load-bearing condition for the central claim is not only that Eq. (5) be accurate; it is that the experiments actually exercise a passive object. Appendix C shows they do not: the object carries a velocity-keeping MPC that constructs a reference line and tracks it. This means the hardware results cannot distinguish the proposed impact-planning stack from the object's own controller. The impact-model mismatch flagged by the reader is real and is acknowledged in Section IX and Fig. 11, but it is downstream of this confound: if the object is actively controlled, even a perfect impact model would not make the experiments a validation of the passive-object claim. The offline MILP planner, zonotope robustness propagation, and MPC are otherwise plausible, and the availability of code and videos is a plus. Lemma 1 in Section IV-C is asserted without proof and could use a formal justification, but it supports the alternative impact-robust metric and is not the main blocker. A clean test with a genuinely passive object, or at least reporting the object's actuator commands from existing logs, would settle the concern. I keep the reader's CONDITIONAL verdict.","tokens_in":19544,"tokens_out":8923,"duration_ms":95209,"concrete_test":"Re-run the throw-and-catch and Pong experiments with the object's velocity-keeping MPC disabled and a truly passive free-floating body. If the planned robot impacts alone do not move the object to the STL goal region with the reported robustness, the hardware validation does not support the central passive-object claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the transported object be passive and uncontrollable: Eq. (9d) sets the object's control input to zero, and the abstract says impacts are the only way to change an object's velocity. Appendix C states that in the hardware experiments \"we run a velocity-keeping MPC on the object\" that \"keeps position, detects an impact, turns off the controller ... and then draws a linear segment from the current state to a future state using this vector.\" The object is therefore actuated and tracks a reference trajectory between impacts, so the hardware results cannot validate the passive-object claim: the observed object motion may be substantially due to the object's own controller rather than to the planned impacts. No report of object control inputs or accelerations is given, so the contribution of the object's actuation is unknown. This also makes the admitted impact-model mismatch (Section IX) and the large Pong deviation (Fig. 11) ambiguous as evidence about the planner, since the object's MPC was active in those trials.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hierarchical planning and control stack for collaborative transportation of passive free-floating objects in microgravity by teams of robots that interact with the objects only through impacts. The stack consists of an offline MILP planner that maximizes either spatial STL robustness or an impact-robustness metric, an online two-body impact replanner, and an impact-aware MPC for execution. The approach is evaluated in a high-fidelity Gazebo simulator and on an air-bearing freeflyer platform with two robots and one object. The authors also provide code and videos. The central claim is that the full stack synthesizes and executes impact-based transport satisfying STL specifications, validated both in simulation and on hardware.","tokens_in":19771,"tokens_out":2806,"duration_ms":29745,"significance":"If the claims hold, the paper provides a useful, computationally tractable framework for a nontrivial robotics problem: transporting objects without grasping, using impacts as the only interaction. The strengths include a complete offline/online/control pipeline, a clear MILP encoding of STL robustness and impact kinematics, the integration with Bezier-curve trajectory parameterization, and the documented open-source implementation. The impact-robustness formulation with zonotope propagation is an interesting idea for accounting for post-impact velocity uncertainty. However, the hardware validation does not support the central passive-object claim, because the object in the experiments runs a velocity-keeping MPC and is therefore actuated between impacts. This confound, together with the unproved Lemma 1 and the deferred robot-object collision avoidance, means that the paper's strongest claims are currently stronger than the evidence supports.","major_comments":[{"comment":"The hardware validation does not support the central passive-object claim. Appendix C states that in the experiments \"we run a velocity-keeping MPC on the object\" and that the object \"keeps position, detects an impact, turns off the controller ... and then draws a linear segment from the current state to a future state using this vector.\" This means the object is actuated between impacts, contradicting the zero-control assumption in Eq. (9d) and the abstract's assertion that impacts are \"the only way to change an object's velocity.\" No object thrust or acceleration telemetry is reported, so the contribution of the object's own MPC to the observed motion is unknown. The Pong deviation and the impact-model mismatch discussed in Section IX are therefore ambiguous as evidence about the planner, since the object controller was active in those trials. Please either repeat the experiments with a truly passive object, or report the object's control inputs and re-analyze the data to separate the object-MPC contribution, or reduce the strength of the hardware-validation claim to what a passive-object experiment would support.","section":"VIII-B and Appendix C"},{"comment":"Lemma 1 is stated without proof, but it is load-bearing for the impact-robust planner. The lemma claims that if for every vertex of the interval hull zonotope there exists a robot trajectory intersecting that vertex at the final time, then for every point in the original zonotope there exists such a trajectory. This requires a convexity or reachability-preservation property that is not established for the B ezier-trajectory formulation with velocity bounds and collision constraints. As written, the lemma is not obviously true for a non-convex or constrained trajectory set. Please provide a proof or a counterexample, or weaken the claim to a conservative heuristic and adjust the statements in Section IV-C accordingly.","section":"Section IV-C, Lemma 1"},{"comment":"The introduction claims the MILP formulation provides \"global optimality guarantees even in the presence of complex spatio-temporal goals and constraints.\" This is only valid for the abstracted point-mass model with Assumptions 1 and 2, and Section V.4 explicitly defers robot-object collision avoidance to the controller safety layer. Consequently, the solution is globally optimal for a relaxed problem, not for the original problem in Eq. (9). The paper's own Limitations section (Section IX) agrees that the robustness metric \"only relates to an abstraction of the true model-based robustness.\" Please qualify the global-optimality statement in the abstract and introduction to refer to the idealized planning model, and state clearly that robot-object collision avoidance is not part of the planning optimization.","section":"Section I and Section V.4"},{"comment":"The planner only enforces collision avoidance between robots and between objects, and \"defers any collision avoidance between robots and objects to safety constraints in the controller.\" This is a significant gap for the claimed \"full planning and control stack\": the STL specifications in the experiments include wall avoidance for all systems, but there is no planning-level guarantee that a robot and an object do not collide at unintended times. The controller-level collision avoidance is only mentioned as a Control Barrier Function in the obstacle scenario, with no formal guarantee or analysis. Please either integrate robot-object collision avoidance into the planning formulation (even conservatively) or explicitly state that unintended robot-object collisions are only filtered by the controller and are not covered by the planner's guarantees.","section":"Section V.4"}],"minor_comments":[{"comment":"The text says the desired pre-impact state is obtained \"from solving Eq. (7) (see Appendix. A)\" but the two-body impact problem is Eq. (6), and Eq. (7) is only the coordinate rotation. Please correct the cross-reference.","section":"Section VI"},{"comment":"Equation (26) states the collision-avoidance constraint with \"||pSi(t)-pSj(t)||2≤ radSi + radSj\" but for avoidance the inequality should be ≥. Please fix the inequality and the corresponding textual explanation.","section":"Section V.4, Eq. (26)"},{"comment":"The footnote says \"the restitution coefficient, c, is 1\" but the rest of the paper uses e for the coefficient of restitution. Please unify the notation.","section":"Section V.4, footnote 1"},{"comment":"The sentence \"It is ˙x+_O,i which is one the left-hand side of Eq. (7)\" contains a spelling error: \"one\" should be \"on.\" Also, the variable name for the desired post-impact object velocity appears as both ˙x+_O,i and ˙xdes_O,i; please make the notation consistent.","section":"Section VI, after Eq. (29)"},{"comment":"The paper would benefit from a pass to harmonize notation for the number of B ezier curves (N vs N_i) and for the indices j in z(R,O)[i,j] and in the equations of Section V. Some indices are introduced but not clearly defined in the MILP encoding subsection.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid systems contribution with a clear pipeline and reproducible open-source artifacts, and I believe the approach is salvageable. The main risk is overclaiming validation: the hardware experiments use an actively controlled object, so the central passive-transport claim is untested on hardware. The unproved Lemma 1 is also a correctness gap that the authors will need to close. I recommend major revision rather than rejection because the issues are identifiable and fixable within the manuscript's scope—provided the experimental claims are softened or the experiments redone with a passive object. One additional concern for the editor: the paper relies heavily on the same authors' prior work [30] for the STL-encoding machinery; this is reasonable but the novelty with respect to [30] should be made crisper in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this is a serious paper on using impacts to move free-floating objects in microgravity, with a full stack—offline MILP planner, online replanner, and impact-aware MPC—under Signal Temporal Logic specifications. The genuinely new pieces are the MILP encoding of impact kinematics with Bézier curves and the robustness metric δ that maximizes permissible post-impact velocity uncertainty. That is a real contribution, and the paper is unusually candid about its limitations.\n\nWhat it does well: the MILP formulation is clean, the zonotope reachability is a nice addition, and the Gazebo simulations run the actual robot software stack. Comparing the spatial-robustness and impact-robustness planners on two scenarios is informative. Code and videos are provided, which is concrete reproducibility evidence. The related work is fair and positions the work well against air hockey, billiards, and space manipulation.\n\nNow the soft spots, in proportion to how much they matter. The stress-test note about the hardware is correct and it is the biggest issue. In Appendix C, the object is not passive: it runs a velocity-keeping MPC that keeps position, detects impacts, turns off briefly to estimate velocity, then tracks a linear segment. So the hardware experiments cannot validate the central claim that impacts are the only way to change an object's velocity. The object's own thrusters shape its motion between impacts, and no control-input data are reported. The simulator results are more convincing because the object is unactuated there, but the notable Pong deviation comes from hardware with that active object, which makes the failure diagnosis ambiguous. This needs to be addressed honestly—either report object actuation and show it is negligible, or reframe the hardware as a demonstration of the replanner/MPC stack without claiming passive-object validation.\n\nSmaller issues: Lemma 1 is unproved; the global-optimality phrase in the abstract overreaches because the MILP is only a model of the real system; and robot-object collision avoidance is deferred to the controller, so tight scenarios may rely on runtime safety. The Limitations section acknowledges most of this.\n\nWho this is for: anyone working on temporal-logic motion planning or impact-driven control will get value. It deserves a serious referee—I would send it—but with a clear request to fix the experimental framing and the missing proof.","headline":"Novel impact-based planning stack with credible simulation but hardware validation is confounded by an actively controlled object.","tokens_in":20238,"tokens_out":3557,"would_cite":true,"duration_ms":35486,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Free-floating objects in space can be transported by a team of robots using only instantaneous impacts, and this paper shows a planner-replanner-MPC stack that realizes it under timed task specifications.","keywords":["collaborative transportation","impact dynamics","signal temporal logic","mixed-integer programming","model predictive control","free-flyer robots","zonotope reachability","microgravity manipulation"],"falsifier":"Run a set of free-flyer collisions at several approach angles and offsets, and compare each measured post-impact object velocity with the prediction of the linear point-mass law using a single restitution coefficient; if the errors routinely exceed the planned margin $\\delta$, the central claim that the planner's robustness metric captures execution uncertainty is falsified. The paper's own Pong experiment shows deviations large enough that this test is worth running.","tokens_in":19393,"feed_emoji":"🚀","tokens_out":7487,"duration_ms":74354,"temperature":0.7,"pith_summary":"In microgravity, a passive floating object does not need to be grasped or carried: robots can move it by hitting it. This paper tries to establish that the full pipeline for that idea works, from an offline planner that chooses impact times and Bézier trajectories to satisfy a formal timed task specification, through an online replanner that accounts for the finite size of the robots and object, to a model predictive controller that executes the plan on hardware. The planner maximizes either spatial robustness of the signal temporal logic specification or, in an alternative formulation, the largest tolerable error in the post-impact velocity of the object. If correct, the approach offers a scalable, low-contact, gripper-free way to transport objects in space, with global optimality guarantees for the simplified planning model.","feed_headline":"Impact-only transport moves floating objects through space","feed_subtitle":"Robots bounce a floating object between them to carry it without grasping, using timed plans that ran on real free-flyers.","key_machinery":"The load-bearing construction is the Bézier trajectory representation in which each physical trajectory $p(t)$ is written as $r(s) = p(h(s))$, a spatial Bézier curve reparameterized by a temporal Bézier curve, with impacts allowed only at curve endpoints. Around this sits the point-mass impact law, which gives linear relations between pre- and post-impact velocities; binary variables select impact versus smooth continuation; zonotopes, convex sets built as Minkowski sums of line segments, propagate the post-impact uncertainty set between impacts; and the impact-robustness margin $\\delta$ measures how much the object's post-impact velocity may deviate before the task fails. Keeping all planning constraints linear is what allows the offline problem to be solved globally as a mixed-integer linear program.","core_discovery":"The central claim is that instantaneous impacts can be the only mechanism for changing a free-floating object's velocity, and that this interaction is simple enough to plan globally. Treating robots and objects as point masses, the planner embeds the linear restitution law as equality constraints on Bézier curve endpoints, with binary variables deciding whether each curve boundary is an impact or a smooth join; the resulting mixed-integer linear program maximizes either STL spatial robustness or an impact-robustness margin $\\delta$. The online replanner relaxes the point-mass assumption by solving the two-body impact geometry with the physical radii, and the MPC tracks the curves while switching its weights near an impact so that the robot's priority is generating the object's desired post-impact velocity. The whole stack is demonstrated in a high-fidelity simulator and on a free-flyer platform for two-robot, one-object scenarios.","pith_inferences":["If the impact model were extended with online estimation of the restitution coefficient and rotational coupling, the same planning pipeline could become reliable enough for on-orbit servicing tasks such as nudging debris or repositioning modules.","The uncertainty-margin formulation suggests a design principle beyond space: in any low-friction environment, choose impact sequences that maximize the allowable error in the impulse, which could transfer to air-hockey or billiards-style manipulation.","A natural testable extension is to make the replanner consider the STL specification, which the paper notes it currently does not; this would trade some online computation for a guarantee that the updated plan remains satisfying rather than merely tracking the nearest impact.","The constant-restitution assumption is the most promising place to add sensing: measuring the object's post-impact velocity after each collision and updating the model parameter would close the loop the current stack leaves open."],"forward_implications":["A team of free-floating robots can transport a passive object without any physical connection, using only bounces, removing the need for grippers or docking mechanisms for such tasks.","Because the offline problem is a mixed-integer linear program, the planner returns globally optimal impact schedules for the simplified model rather than locally adjusted ones.","The robustness variant suggests that extra impacts can shrink rather than grow the object's uncertainty funnel, since each impact can be planned to correct velocity errors.","The same stack of offline planner, online replanner, and MPC applies to any timed task expressible as STL reach-avoid constraints, not only point-to-point transport.","Simulation and hardware runs show that the approach is executable in real time on free-flyer platforms, subject to the idealized impact model."],"supporting_citations":[{"why":"Supplies the Bézier-curve convex-optimization trajectory formulation that the planner builds on for smooth, collision-free curves.","marker":"[17]"},{"why":"Extends the Bézier formulation to multi-agent planning with STL specifications, providing the robustness embedding used here.","marker":"[30]"},{"why":"Gives the piecewise-linear trajectory STL robustness embedding that the offline planner relies on for object trajectories.","marker":"[27]"},{"why":"Defines STL and its spatial robustness semantics, which are the objective of the spatially robust planner.","marker":"[16]"},{"why":"Supplies the zonotope reachability properties used to propagate post-impact velocity uncertainty between impacts.","marker":"[7]"},{"why":"Introduced embedding of STL specifications in mixed-integer optimization problems, the basis for the planning constraints.","marker":"[23]"},{"why":"Describes the free-flyer experimental platform used for hardware validation.","marker":"[24]"},{"why":"Is the mixed-integer solver used to compute the offline MILP plans.","marker":"[8]"},{"why":"Is the optimization framework used for the online replanner and the model predictive controller.","marker":"[2]"}],"fun_headline_variants":["Impact-only method lets robots ferry objects in space","Bounce planning: robots move floating objects without grasp","Zero-g transport via timed impact interactions","Robots use impacts to carry objects in microgravity","Impact-driven object transport tested on free-flyers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The plan assumes impacts follow an idealized point-mass collision law with a fixed restitution coefficient and no rotation; if real collisions transfer spin or depend on shape and contact angle, the planned guarantees and the uncertainty margin $\\delta$ are only approximations of what happens physically.","fun_headline_variants_meta":{"raw":{"variants":["Impact-only method lets robots ferry objects in space","Bounce planning: robots move floating objects without grasp","Zero-g transport via timed impact interactions","Robots use impacts to carry objects in microgravity","Impact-driven object transport tested on free-flyers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000507,"raw_usage":{"total_tokens":2446,"prompt_tokens":896,"completion_tokens":1550,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":1478}},"tokens_in":512,"tokens_out":1550,"duration_ms":9385,"temperature":1.0,"reasoning_tokens":1478,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:12:19.253323+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a set of free-flyer collisions at several approach angles and offsets, and compare each measured post-impact object velocity with the prediction of the linear point-mass law using a single restitution coefficient; if the errors routinely exceed the planned margin $\\delta$, the central claim that the planner's robustness metric captures execution uncertainty is falsified. The paper's own Pong experiment shows deviations large enough that this test is worth running.","supporting_citations":[{"cited_title":"Temporally robust multi-agent stl motion planning in continuous time","cited_arxiv_id":null,"evidence_quote":"Extends the Bézier formulation to multi-agent planning with STL specifications, providing the robustness embedding used here."},{"cited_title":"Multi-agent motion planning from signal temporal logic specifications","cited_arxiv_id":null,"evidence_quote":"Gives the piecewise-linear trajectory STL robustness embedding that the offline planner relies on for object trajectories."},{"cited_title":"Monitoring temporal properties of continuous signals","cited_arxiv_id":null,"evidence_quote":"Defines STL and its spatial robustness semantics, which are the objective of the spatially robust planner."},{"cited_title":"Reachability of uncertain linear sys- tems using zonotopes","cited_arxiv_id":null,"evidence_quote":"Supplies the zonotope reachability properties used to propagate post-impact velocity uncertainty between impacts."},{"cited_title":"Model predictive control from signal temporal logic specifications: A case study","cited_arxiv_id":null,"evidence_quote":"Introduced embedding of STL specifications in mixed-integer optimization problems, the basis for the planning constraints."},{"cited_title":"CasADi: a software framework for nonlinear optimization and optimal con- trol","cited_arxiv_id":null,"evidence_quote":"Is the optimization framework used for the online replanner and the model predictive controller."}],"review_version":1}