{"id":"fea37a07-112e-4892-9df2-ebfa92a37f66","arxiv_id":"2501.02815","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A spatial-domain reactive controller that propagates control along the robot's kinematic chain and enforces link-specific collision constraints shows better task completion in cluttered mobile manipulation tasks.","lead":"This paper describes a new reactive controller for mobile manipulators that plans over the robot's kinematic chain rather than over time, allowing each arm link to be checked for collisions separately. The authors report improved success rates and shorter paths in cluttered simulations and on a real robot, but key evidence such as computation time is missing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-time performance is asserted but never measured: no solve times or control periods are reported, so the central 'real-time reactive controller' claim is unsupported.","rationale":"The reader's verdict is CONDITIONAL with moderate confidence, and my read agrees with that verdict. I do not see an internal inconsistency that warrants rejection; the spatial-domain reformulation is coherent and the reported success rates are plausible. The most load-bearing weakness is the complete absence of timing evidence for the 'real-time' assertion, because the contribution is defined in opposition to finite-horizon MPC and single-step QP controllers on computational grounds. The free-region/SOS safety assumption is also important and is well captured by the reader, but it is inherited from prior work and is partially supported by the obstacle-free simulation results; the timing claim, by contrast, has no quantitative support at all. A single instrumentation pass can settle it. Thus the conditional verdict should stand, with a required revision to add solve-time statistics and control periods.","tokens_in":10784,"tokens_out":9206,"duration_ms":92329,"concrete_test":"Instrument the released code in the cluttered-forest scenario (density 0.7 obstacles/m2, 3 start-goal pairs, 20 repetitions each) to record wall-clock solve time per control cycle for the proposed AL-DDP/SOS pipeline; report mean, p95, and max over at least 60 runs, together with the controller's period. Also record the same statistic for SLQ-MPC under identical hardware. If p95 solve time is below the control period (e.g., 10 Hz implies 100 ms), the real-time claim is supported; if p95 exceeds it, the reactive-controller and efficiency claims fail. This check is minimal and directly addresses the gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central value proposition is a real-time local reactive controller: the abstract claims 'efficiency', the introduction and conclusion assert 'real-time', and Tables I–II include a 'Real-time' column. Yet no solve time, control frequency, or computational budget is reported anywhere. Section IV states only the hardware (Intel i7-13300H laptop for simulation; i5-8600 on the robot) and says the problem is solved with ALTRO plus SOS/SDP subproblems via COPT. The 'Real-time ✓' entries have no threshold, so they cannot be checked. This matters because AL-DDP with per-link SOS containment subproblems is computationally heavy; the authors themselves exclude NEO from the forest experiment due to 'extended solving time' (Section IV.A.1), showing that solve time is a real differentiator. If per-step solve time exceeds the control period, the method is not a reactive controller and the claimed efficiency advantage collapses. The reader's weakest_assumption focuses on the free-region/SOS safety constraint, which is a valid conditional concern, but the timing gap is more directly load-bearing for the central claim: without timing, the 'real-time' part of the claim is unfalsified and the comparison against MPC/QP baselines is incomplete.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a local reactive controller for mobile manipulators that reformulates the usual time-domain single-step reactive control problem as a multi-step optimization problem in the spatial domain along the kinematic chain. The state of each link is optimized over a spatial horizon using spatial kinematic propagation functions, link-specific costs, and collision-avoidance constraints defined by containment of each link's polytopic geometric model in a free region. The constraints are enforced via an SOS-derived minimum scaling factor, and the resulting problem is solved with AL-DDP (ALTRO). The authors evaluate the method in a simulated cluttered forest, a combined navigation-manipulation task, and real-world experiments with a mobile base and 6-DOF Kinova arm, comparing against RRT, SLQ-MPC, Coupled-MPC, and NEO. They report higher success rates and shorter path lengths for their method in the multifaceted task and a comparable success rate in the forest, along with real-world demonstrations.","tokens_in":11012,"tokens_out":6283,"duration_ms":58463,"significance":"The central idea—moving from a single-step time-domain reaction to a multi-step spatial-domain optimization along the serial chain—is a genuinely new way to decouple whole-body constraints in reactive control, and it is supported by a plausible forward-kinematics formulation for an omnidirectional base. The link-specific free-region/SOS collision representation is a promising alternative to sphere-based or ESDF-based safety constraints, and the authors provide an open-source repository, which increases reproducibility. If the real-time and robustness claims are substantiated with timing measurements and statistical reporting, the contribution would be useful to the mobile-manipulation community. At present, however, the paper's own evidence is incomplete: no solve-time data are given, safety is conditional on inherited SOS/free-region machinery, and Tables I–II report aggregate point estimates without variance.","major_comments":[{"comment":"The paper's central claim is that the controller is real-time and efficient, but no timing measurement appears anywhere: no solve time, control period, or computation budget is reported, and the 'Real-time ✓' entries in Tables I and II have no threshold. This is load-bearing because Section IV.A.1 explicitly excludes NEO from the forest experiment due to 'extended solving time,' showing that solve time is a differentiator. Please report per-step and per-iteration solve times, control frequency, and hardware, including timing for failed or degraded runs, so the real-time claim can be checked.","section":"Section IV, Tables I–II"},{"comment":"The whole-body safety guarantee is inherited from the free-region decomposition [15] and the SOS containment subproblem [18], and the paper does not analyze failure modes of those components (e.g., poor free regions, nonsmooth alpha_k gradients, or aggressive solutions with minimal clearance). The paper itself reports 'minor scratches' in failure cases in Section IV.A.1, which weakens any unconditional safety claim. Please state the precise safety property being guaranteed, analyze when the SOS/free-region constraints can fail, and report collision/clearance statistics rather than only success rates.","section":"Section III-D, Eqs. (3)–(4)"},{"comment":"The experimental tables report success rates and path lengths over 20 repetitions with no error bars, standard deviations, or statistical tests, and NEO is omitted from the forest comparison with 'N/A' entries. Because success rates such as 0.9 versus 0.9 and path-length differences around 1–2 m may be within noise, these point estimates do not support the claimed 'significant improvements.' Report per-path or per-repetition distributions, define the success criterion explicitly, and include NEO's timing or justify why its omission does not bias the forest comparison.","section":"Tables I–II, Section IV.A.1"},{"comment":"The spatial reformulation, which is the core contribution, would benefit from a more precise statement of decision variables and the receding-horizon update. Equation (1) minimizes over Q and U, but the constraints (1a)–(1d) are written for q^{t+1}_k with fixed t, and the relationship between the spatial sequence {q_k, u_k} and the next time step q^{t+1} is not explicit. Please clarify the horizon N̄, the link indexing for base and arm, and how the optimized sequence is applied at each control step.","section":"Section III-A, Eqs. (1)–(2)"}],"minor_comments":[{"comment":"There are typos such as 'challanges' in the introduction, and the two consecutive paragraphs beginning 'On the other hand' should be restructured for readability.","section":"Section I"},{"comment":"The table captions should define the columns 'Real-time' and 'Reference Requirement,' and state the number of repetitions and the success criterion used to compute the success rate.","section":"Tables I–II"},{"comment":"The text says '20 repetitions for each path,' but the table aggregates over three start-goal pairs; reporting per-path results would make the effect of obstacle density and initial path quality visible.","section":"Section IV.A.1"},{"comment":"The bounding-box limits for free-region generation (0.8 m x 0.8 m x 0.8 m for the base and 0.4 m x 0.3 m x 0.3 m for links) are introduced without explaining how the bounding boxes are incorporated into the decomposition; a sentence or figure reference would help.","section":"Section III-D"},{"comment":"The open-source repository link appears only in the abstract; a Data Availability or Code Availability statement should be added in the main text.","section":"Abstract and Section IV"},{"comment":"The phrase 'The original points of the virtual links q1, q2, and q3 are located at the center of the mobile base' should be 'origins,' and the notation for the base control inputs (v_base_x, v_base_y, phi_dot) should be introduced explicitly before Eq. (2).","section":"Section III-C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope and the spatial-domain formulation is promising, but I would not accept it without timing measurements and a statistically strengthened experimental section. Please also ask the authors to clearly delineate what is new relative to their prior work [15] and [18], since the safety-constraint machinery is largely inherited. The omitted NEO comparison in the forest experiment should be addressed rather than hidden behind N/A. No ethical or attribution concerns beyond the above."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is worth your time. Reformulating the single-step time-domain reactive control problem as a multi-step optimization along the kinematic chain is a real departure from the cited literature. The link-wise decoupled constraints, built on SOS-based free-region containment, are a sensible extension of the authors' prior work, and the experiments cover a plausible range of baselines (RRT, SLQ-MPC, Coupled-MPC, NEO) in simulation plus a real-world demo. Credit where due: the forward-kinematics propagation is clearly explained, the math is coherent, and they ship code.\n\nThe biggest problem is what the stress-test note flags: 'real-time' is asserted everywhere but never measured. No solve time, control frequency, or computational budget appears anywhere. Tables I and II have a 'Real-time' column with checkmarks, but there is no threshold or measurement. For a method billed as a reactive controller, this is load-bearing. The authors even exclude NEO from the forest experiment because of 'extended solving time,' which shows timing is a differentiator, yet they do not report their own numbers. That needs to be fixed with actual wall-clock data before the central claim can be trusted.\n\nOther soft spots: the 'whole-body safety' claim is stronger than the evidence. They admit failure cases with 'minor scratches' in the forest experiment, which is understandable, but it means the safety result is empirical, not a formal guarantee. Also, the tables lack error bars, and some algorithmic details are deferred to [18], making independent verification harder. The omission of NEO in the forest baseline is unexplained beyond a note; if NEO was too slow, that is fine, but report its timing.\n\nBut keep this proportional: the core idea is not circular, the evaluation uses external baselines, and the math is plausible. This paper is above the desk-reject line.\n\nWho it is for: researchers working on reactive control or whole-body motion planning for mobile manipulators. A serious referee should engage, mainly to demand timing data and error bars. If the authors can show solve times under, say, 10-20 ms per control step on the stated hardware, the real-time claim becomes credible. I would send it to peer review with a request for revision.","headline":"A genuinely new spatial-domain reformulation of reactive control for mobile manipulators, worth a close look, but the central real-time claim is unsupported by any timing data.","tokens_in":11567,"tokens_out":1658,"would_cite":true,"duration_ms":16419,"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":"A mobile manipulator can be controlled with whole-body safety in real time by recasting reactive control as a multi-step optimization along the kinematic chain, with per-link geometric collision constraints solved by AL-DDP.","keywords":["mobile manipulators","reactive control","whole-body safety","collision avoidance","spatial trajectory optimization","augmented Lagrangian DDP","sums-of-squares programming","free regions"],"falsifier":"Run the controller toward a thin vertical pole placed between two consecutive links so that no single link's body enters the pole, while the pole still lies inside the 3 m perception range. If the optimizer reports $\\alpha_k \\le 1$ for every link and the pole nevertheless contacts the robot because the adjacent free regions leave an uncovered gap, the whole-body safety claim fails.","tokens_in":10565,"feed_emoji":"🤖","tokens_out":7626,"duration_ms":71182,"temperature":0.7,"pith_summary":"Mobile manipulators are hard to control reactively because every link's motion is coupled to all preceding joints, so whole-body collision safety is usually approximated with spheres or ignored. This paper tries to establish that the coupling is an artifact of solving the control problem in the time domain, and it reformulates the single-step reactive problem as a multi-step optimization along the spatial kinematic chain, treating each link as one step. In that form, each link gets its own constraints, including a geometry-accurate collision constraint that requires the link to stay inside a free region extracted from the local map. The resulting optimization is solved with augmented Lagrangian differential dynamic programming (AL-DDP), which propagates the spatial kinematics in the forward pass and handles all link-specific constraints simultaneously in the backward pass. If the method works as reported, high-degree-of-freedom articulated robots can react to unknown obstacles in real time while preserving whole-body safety and threading narrow gaps.","feed_headline":"Spatial reformulation makes whole-body robot safety real-time","feed_subtitle":"Optimizing along the kinematic chain instead of over time clears cluttered paths with precise link geometry.","key_machinery":"The central object is the spatial multi-step trajectory optimization of Eq. (1), where the state of each link on the serial chain is $q_k = [p_k, r_k]$ and the objective runs over the chain instead of over time, with kinematic propagation $f_k$ linking consecutive states. The mechanism that carries the argument is the pairing of link-specific geometric containment constraints $W_k \\subseteq F_k$ with AL-DDP, an optimal-control solver that pulls constraints into the cost and propagates dynamics forward and backward. The collision constraint is made differentiable by a sums-of-squares subproblem that outputs the minimum scaling factor $\\alpha_k$ and its gradients, and the safety condition is $\\alpha_k \\le 1$ for every link. This machinery turns whole-body safety into a set of decoupled per-link constraints that can be evaluated simultaneously during the backward pass.","core_discovery":"The paper's central claim is that reformulating the time-domain single-step reactive problem into a multi-step spatial trajectory optimization along the kinematic chain is what makes whole-body safety tractable. Each link becomes a step with its own state $q_k$ and control $u_k$, and the propagation $f_k$ between adjacent links is absorbed naturally in the solver's forward pass. Collision avoidance is expressed per link as a containment constraint $W_k \\subseteq F_k$ between the link's exact geometric body and a polytopic free region, enforced by a sums-of-squares computed minimum scaling factor $\\alpha_k \\le 1$ with gradient information. The decoupled constraint structure lets the backward pass of AL-DDP process all link-specific constraints simultaneously, yielding real-time computation with improved safety, shorter paths, and higher task completion rates than the baselines tested.","pith_inferences":["(Editorial inference) The spatial horizon idea is a general recipe for serial kinematic chains, so humanoid limbs, dual-arm systems, or snake-like robots could inherit the same decoupling, provided free regions can be generated along their skeletons.","(Editorial inference) The safety certificate is local: it covers only what the 3 m perception range sees and only when each link is separately contained in its own free region, so a fully safe guarantee under arbitrary unknown obstacles would need an argument about gaps between adjacent free regions.","(Editorial inference) A natural stress test is to move a thin obstacle into the space between two consecutive links; if the per-link scaling-factor constraints all stay below one and the robot still contacts the obstacle, the whole-body claim would need an inter-link condition in addition to the per-link one."],"forward_implications":["Mobile manipulators can be driven by a local reactive controller that checks every link's exact geometry against locally perceived obstacles at real-time rates, without waiting for a global replan.","Because each link is constrained separately, adding a new per-link objective or constraint does not entangle it with the other links' constraints, so task-specific costs, such as keeping the base or a chosen arm link on a desired path, become straightforward.","The reported experiments indicate higher task completion rates and shorter traveled paths in narrow, cluttered scenes compared with MPC, QP, and sampling baselines, while remaining online with only a rough global path as guidance.","The approach extends the known free-region plus sums-of-squares collision formulation from a single rigid body to a whole articulated robot, which is the step that lets the robot use its actual link shapes in tight spaces."],"supporting_citations":[{"why":"Finite-horizon MPC baseline using an ESDF for safety constraints; the method must beat it in real-time reactivity and conservativeness.","marker":"[1]"},{"why":"Single-step QP reactive controller for manipulators used as a baseline showing the failure mode of joint-space velocity-limited control.","marker":"[4]"},{"why":"Free-region decomposition algorithm that produces the polytopic free regions along the robot's skeleton used in every collision constraint.","marker":"[15]"},{"why":"Coupled MPC baseline with free-space decomposition and plane-distance collision constraints, the prior art that the paper's per-link decoupling improves on.","marker":"[16]"},{"why":"Sums-of-squares subproblem that computes the minimum scaling factor alpha_k and its gradients for the containment condition W_k subset F_k.","marker":"[18]"},{"why":"Solver providing the augmented Lagrangian DDP machinery used to solve the spatial trajectory optimization problem.","marker":"[22]"}],"fun_headline_variants":["Spatial reformulation achieves real-time whole-body robot safety","Whole-body safety in real time via spatial trajectory optimization","Per-link constraints make mobile manipulator safety real-time","Spatial-domain control clears cluttered paths with whole-body safety","AL-DDP with per-link geometry makes robot safety real-time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The safety claim rests on the assumption that each link can be enclosed in a free region generated along the robot's skeleton and that the sums-of-squares minimum scaling factor $\\alpha_k \\le 1$ is a reliable test for collision with everything inside the 3 m perception range.","fun_headline_variants_meta":{"raw":{"variants":["Spatial reformulation achieves real-time whole-body robot safety","Whole-body safety in real time via spatial trajectory optimization","Per-link constraints make mobile manipulator safety real-time","Spatial-domain control clears cluttered paths with whole-body safety","AL-DDP with per-link geometry makes robot safety real-time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000486,"raw_usage":{"total_tokens":2397,"prompt_tokens":946,"completion_tokens":1451,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":1378}},"tokens_in":562,"tokens_out":1451,"duration_ms":10635,"temperature":1.0,"reasoning_tokens":1378,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:03:32.079682+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the controller toward a thin vertical pole placed between two consecutive links so that no single link's body enters the pole, while the pole still lies inside the 3 m perception range. If the optimizer reports $\\alpha_k \\le 1$ for every link and the pole nevertheless contacts the robot because the adjacent free regions leave an uncovered gap, the whole-body safety claim fails.","supporting_citations":[{"cited_title":"Perceptive model predictive control for continuous mobile manipulation,","cited_arxiv_id":null,"evidence_quote":"Finite-horizon MPC baseline using an ESDF for safety constraints; the method must beat it in real-time reactivity and conservativeness."},{"cited_title":"NEO: A novel expeditious optimisation algorithm for reactive motion control of manipulators,","cited_arxiv_id":null,"evidence_quote":"Single-step QP reactive controller for manipulators used as a baseline showing the failure mode of joint-space velocity-limited control."},{"cited_title":"Planning dynamically feasible trajectories for quadrotors using safe flight corridors in 3-d complex environments,","cited_arxiv_id":null,"evidence_quote":"Free-region decomposition algorithm that produces the polytopic free regions along the robot's skeleton used in every collision constraint."},{"cited_title":"Coupled mobile manip- ulation via trajectory optimization with free space decomposition,","cited_arxiv_id":null,"evidence_quote":"Coupled MPC baseline with free-space decomposition and plane-distance collision constraints, the prior art that the paper's per-link decoupling improves on."},{"cited_title":"Collision-free trajectory optimization in cluttered environments using sums-of-squares programming,","cited_arxiv_id":null,"evidence_quote":"Sums-of-squares subproblem that computes the minimum scaling factor alpha_k and its gradients for the containment condition W_k subset F_k."},{"cited_title":"ALTRO: A fast solver for constrained trajectory optimization,","cited_arxiv_id":null,"evidence_quote":"Solver providing the augmented Lagrangian DDP machinery used to solve the spatial trajectory optimization problem."}],"review_version":1}