{"id":"479c0244-f1fb-4ae7-b40d-274ccf19d5cb","arxiv_id":"2607.17007","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A distributed model-predictive controller with ADMM coordination lets teams of quadrupedal robots safely carry a shared payload around obstacles, with simulations and hardware demos.","lead":"Robots that walk on four legs can carry a shared box together while avoiding obstacles, using a distributed controller that splits planning into per-robot pieces coordinated by ADMM. The paper reports simulations and real experiments with two- and three-robot teams, claiming up to 23% faster planning than a centralized controller with similar performance.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Safety-critical claim is undermined by the paper's own reported HOCBF boundary violations; no full-order tracking or independent ground-truth analysis rules out real violations.","rationale":"The reader's weakest assumption already identifies the same load-bearing concern: safety depends on HOCBFs applied to a reduced-order SRB model with MoCap-based state estimates, while the experiments show apparent safety-boundary violations. This is precisely the weakest link in the central claim. I checked the discrete-time HOCBF condition in Eq. (21) and it is consistent with the standard forward-difference formulation from [33], so the mathematical form of the safety constraint is not the problem. The unresolved issue is the gap between the SRB-level planned safety and the full-order tracked behavior, which the paper's own experiments expose. The solve-time and yaw-tracking improvements are credible and independently supported by ablation trends, but they do not require the 'safety-critical' label to be true. The CONDITIONAL verdict remains appropriate: a revision that either drops the safety-critical guarantee, provides full-order ground-truth distance measurements, or supplies a formal tracking-error bound would address the concern. No verdict change is needed beyond the reader's existing conditional assessment.","tokens_in":13512,"tokens_out":7939,"duration_ms":72915,"concrete_test":"Re-analyze the logged MoCap, joint-encoder, and coupling kinematics from the nominal and push experiments at the exact time instants where h crosses the zero boundary in Figs. 3-5. Reconstruct the full-order robot/payload geometry (body, limbs, and payload basket) from the measured joint angles and rigid-coupling model, and recompute the minimum Euclidean distance from the true full-order volumes to each obstacle. If any distance falls below the physical clearance corresponding to d_safe = 0.6 m, the HOCBF violation is real and the safety-critical claim is falsified. If all full-order distances remain above d_safe, the 'apparent' violation is confirmed as a state-estimation artifact, resolving the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central 'safety-critical' claim rests on the assumption that the HOCBF constraints (7)-(8), computed on the reduced-order SRB model and enforced in each local NMPC (17), imply actual collision avoidance for the full 18-DoF robots once tracked by the decentralized WBC. The paper's own hardware results do not establish this. Section IV-B reports 'a slight apparent violation of two HOCBF safety functions' in nominal two-agent experiments and 'strong pushes may induce brief apparent safety-boundary violations' in robustness experiments, with these attributed to MoCap measurement noise rather than real safety breaches. But h in (6) is computed from SRB CoM estimates, so a logged h crossing below zero could equally be a genuine robot/payload entering the 0.6 m safety margin, a WBC tracking error, or a sensor artifact. No formal bound relates SRB-planner feasibility plus WBC tracking error to full-order h >= 0, and no independent ground-truth measurement of the full-order robot or payload geometry is provided. Therefore the paper's headline 'safety-critical' guarantee is not established; the experiments support heuristic safety at best. This concern does not invalidate the solve-time or tracking improvements, but it directly undercuts the strongest advertised claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes an ADMM-based distributed nonlinear model predictive control (NMPC) framework for cooperative payload transportation by teams of quadrupedal robots. The centralized finite-horizon optimal control problem is decomposed into parallel local NMPC subproblems, with consensus enforced over both payload-state and interaction-wrench trajectories. Each local problem incorporates single-rigid-body dynamics, acceleration-level holonomic coupling constraints, and higher-order control barrier function (HOCBF) safety constraints. The framework is evaluated in simulation with two, three, and four Unitree Go2 agents and in hardware experiments with two-agent and three-agent (heterogeneous Go2/A1) teams, under payload uncertainty, external disturbances, and various obstacle configurations. The reported results include up to 23% lower average NLP solve time than centralized NMPC for three- and four-agent cases, and an ablation indicating that payload-state consensus with holonomic constraints reduces payload yaw RMS error by approximately 4x relative to a wrench-only approximation.","tokens_in":13829,"tokens_out":6829,"duration_ms":64652,"significance":"If the central claims are substantiated, this would be a useful contribution to cooperative legged manipulation: it extends wrench-only ADMM formulations to include payload-state consensus and explicit holonomic constraints, and it provides unusually detailed hardware validation and ablation studies. The solve-time comparison is based on measured NLP solver times, the yaw-error improvement is a measured ablation outcome, and the paper reports ADMM residual behavior and parameter sensitivity, which are commendable. However, the headline 'safety-critical' property is not established by the evidence as presented: the paper's own hardware logs show HOCBF boundary violations that are attributed to measurement noise without independent verification, and no formal or experimental ground-truth link is provided between reduced-order planner safety and full-order robot safety. The contribution is therefore conditionally significant: it demonstrates a promising and well-engineered distributed optimization architecture, but the strongest advertised claim needs additional support.","major_comments":[{"comment":"The 'safety-critical' claim is undermined by the paper's own reported data. The text states that a 'slight apparent violation of two HOCBF safety functions' occurs in nominal experiments and that 'strong pushes may induce brief apparent safety-boundary violations,' with all such events attributed to MoCap measurement noise. Since the safety function h in (6) is computed from estimated SRB center-of-mass states, a logged h<0 could equally indicate a genuine breach of the 0.6 m safety margin, a whole-body-control tracking error, or a sensor artifact. No formal bound relates SRB-planner feasibility plus WBC tracking error to full-order h>=0, and no independent ground-truth measurement of the full robot/payload geometry is provided. The title, abstract, and contribution statement in Section I-B therefore claim more than the evidence supports. Please either provide a robustness bound or indep","section":"Section IV-B, Figs. 3-5; Section I-B"},{"comment":"The acceleration-level holonomic constraint is stated but not derived. Eq. (4) defines the holonomic constraint phi_hol(x_i, x_L)=0, and Eq. (5) asserts that double differentiation yields phi_hol_ddot(x_i, x_L, u_i, lambda_L)=0. Because x_i includes Euler angles and the dynamics are discrete-time, the explicit expression is nontrivial. This constraint is load-bearing: it is enforced in each local feasible set Z_i and used in the local NMPC (17). Without the explicit derivation, or at least a precise statement of how phi_hol_ddot is computed (symbolic differentiation, finite differences, etc.), the formulation is not fully reproducible and the exact nature of the coupling constraint is unclear.","section":"Section II-B, Eq. (5)"},{"comment":"The comparison against 'wrench-only ADMM formulations' is not a faithful comparison. The ablation reduces rho_x and removes the holonomic constraints, which is a degraded version of the proposed architecture, not the algorithm of [28]. Moreover, the ablation changes two features at once, so the reported 4x reduction in RMS payload yaw error cannot be attributed specifically to payload-state consensus or to holonomic constraints. To support the contribution claim, implement a faithful wrench-only baseline or separate the two ablations.","section":"Section IV-C2, Table II"},{"comment":"The statement that the distributed framework achieves 'comparable closed-loop performance' to centralized NMPC is not supported by any quantitative data. Table I reports only NLP solve times; no closed-loop tracking errors, obstacle clearance margins, or payload orientation errors are compared between distributed and centralized formulations. Without such a comparison, the reader cannot evaluate whether the computational gains come at a performance cost. Please add closed-loop performance metrics for both formulations.","section":"Section IV-A, Table I"}],"minor_comments":[{"comment":"The five-dimensional holonomic constraint (three translational, two rotational) is described verbally but the actual constraint function for the rigid coupling mechanism and the yaw-free joint is not written out. Please provide the explicit expression.","section":"Section II-B"},{"comment":"A terminal weighting matrix P_SRB = 10 Q_SRB is mentioned, but the cost (10) has no terminal cost term. Clarify how P_SRB enters the local NMPC (e.g., as a terminal state penalty appended to the horizon).","section":"Section IV-A"},{"comment":"The column labeled 'Iterations' is ambiguous: it could mean IPOPT iterations or ADMM iterations. Please clarify the notation.","section":"Table I"},{"comment":"The residual plots labeled 'RMS(m)' and 'RMS(N)' should state how the RMS is computed over the prediction horizon and over time, and whether it is an average over agents.","section":"Figs. 3-5"},{"comment":"The hardware implementation uses one offboard computer with multiple threads; the 'distributed' property is algorithmic, not network-based. This is acknowledged, but it is worth restating that the communication-delay ablation is an emulation.","section":"Section IV-A"},{"comment":"Several key comparisons and prior works are arXiv preprints (e.g., [24], [28]). Please indicate in the text which results are published and which are preprint, and whether the comparison implementation is publicly available.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The experimental effort is substantial and the optimization architecture is coherent. The main concern is the mismatch between the 'safety-critical' branding and the evidence, which I believe is addressable with additional validation or more careful claims. I also note the heavy reliance on the authors' own recent preprints; this is acceptable for a research sequence, but the comparison against [28] should be made faithful. The paper is within the journal's scope and the central derivation is sound, so major revision rather than rejection is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper before reading it. First, the authors actually built and tested this: two- and three-agent quadruped teams carrying a physical payload through cluttered spaces, with an ADMM-based distributed NMPC that enforces consensus over both payload-state and interaction-wrench trajectories. That is a real step beyond the wrench-only and centralized comparisons. Second, the headline \"safety-critical\" is not established by the evidence they present. The paper itself reports HOCBF safety functions dipping below zero in both nominal and pushed experiments, and attributes this to MoCap noise without independent ground truth. Since h is computed from SRB state estimates, a logged violation could be a real geometry intrusion, a WBC tracking error, or sensor artifact. No formal bound relates the reduced-order planner plus WBC to full-order safety, so the experiments support heuristic safety, not a guarantee. That undercuts the strongest advertised claim, but not the distributed optimization contribution. What is actually new: including holonomic coupling constraints and payload-state copies inside each local NLP, with parallel IPOPT solves and a fixed small number of ADMM iterations. The solve-time reduction (9–23% for 3–4 agents) is modest but real, and the 4x yaw-error improvement over the reduced-gain wrench-only ablation is meaningful. The formulation is standard ADMM/NMPC/HOCBF building blocks, cleanly assembled. Soft spots, in proportion: (i) the acceleration-level holonomic constraint (5) is stated but never derived; you need that derivation to verify the claimed dynamics-consistency improvement. (ii) The wrench-only baseline is approximated by lowering rho_x and dropping holonomic constraints, not the actual algorithm from [28]; fine for intuition, weak as a comparison. (iii) No code or data. Those are all fixable in revision. The central argument for the distributed architecture holds up for what it is: a well-engineered, parallelizable controller with credible hardware results. The safety-critical label should be toned down to \"safety constraints enforced in the planner\" unless they close the gap between SRB feasibility and full-order safety. My bottom line: send it to peer review, but require the holonomic derivation, a ground-truth or bounding argument for the HOCBF violations, and a real baseline comparison. The robotics community will get value from the architecture and the experiments even if the guarantee language is softened.","headline":"A useful, hardware-validated distributed MPC for quadruped payload transport, but the safety-critical claim is softer than advertised; worth a serious referee with major revisions.","tokens_in":14360,"tokens_out":993,"would_cite":true,"duration_ms":12226,"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 distributed predictive controller lets teams of quadrupedal robots carry a shared payload through cluttered spaces, coordinating both payload state and interaction forces while reducing computation time.","keywords":["distributed model predictive control","ADMM","quadrupedal robots","cooperative payload transportation","control barrier functions","holonomic constraints","safety-critical control","multi-robot systems"],"falsifier":"Run the two- or three-agent transportation task using onboard state estimation instead of an external motion-capture system, or with a tighter safety margin than the reported threshold, and record the minimum value of the HOCBF safety functions. If any safety function goes negative for longer than a single control step, or the payload yaw error exceeds the roughly 4x margin reported relative to wrench-only consensus, the central safety-critical claim is contradicted.","tokens_in":13390,"feed_emoji":"🤖","tokens_out":4450,"duration_ms":42608,"temperature":0.7,"pith_summary":"The paper tries to establish that a team of load-carrying quadruped robots can be controlled by a distributed nonlinear predictive controller rather than a centralized one, without losing safety or tracking quality. It splits the coupled robot-payload planning problem into parallel local controllers that agree on the shared payload trajectory and the interaction wrenches through ADMM, while enforcing obstacle-avoidance constraints with higher-order control barrier functions. If true, this would make cooperative transportation scalable and real-time for larger teams, since per-robot optimization time grows more slowly than centralized optimization. The paper reports up to 23 percent lower average solve time for four-agent teams, comparable tracking, and about four times lower payload yaw error than a wrench-only consensus baseline.","feed_headline":"Quadruped teams carry payloads with distributed predictive control","feed_subtitle":"A distributed NMPC with ADMM consensus over payload state and wrenches cuts solve time by up to 23 percent vs. centralized.","key_machinery":"The central object is the ADMM-based distributed optimization architecture: each agent maintains local copies of the payload trajectory and all interaction wrenches, penalized toward a global consensus trajectory with a block-diagonal penalty matrix. The local subproblems include the single-rigid-body dynamics, acceleration-level holonomic constraints from rigid coupling, and higher-order control barrier function (HOCBF) constraints for obstacle avoidance. Two ADMM iterations per control update are used, with parallel local solves and warm-starting.","core_discovery":"The central claim is that safety-critical cooperative transportation of a shared payload by multiple quadrupeds can be decomposed into parallel local NMPC subproblems that reach consensus over both the payload-state trajectory and the interaction-wrench trajectories, with acceleration-level holonomic coupling constraints and HOCBF safety constraints included directly in the distributed formulation. Unlike prior wrench-only ADMM formulations, this architecture keeps local copies of the payload state so each robot predicts a dynamically consistent shared trajectory. The authors validate this in simulation for two, three, and four agents and in real-time experiments on two- and three-robot team","pith_inferences":["If the state-consensus advantage holds at larger scales, wrench-only distributed controllers for legged loco-manipulation could be upgraded by adding local payload-state copies rather than a separate payload solver, at modest communication cost.","The safety-critical claim currently rests on motion-capture state and a reduced-order model; a natural test is whether the same HOCBF layer preserves safety under onboard-only state estimation or model error in the whole-body tracking layer.","Because the paper notes formal convergence for finite-iteration nonconvex ADMM remains open, robustness in practice depends on the fixed two-iteration schedule; an extension could adapt the number of iterations online based on residual thresholds."],"forward_implications":["Three- and four-agent transportation can be solved in real time with about 9% and 23% lower average NLP solve time than centralized NMPC while keeping similar closed-loop performance.","Explicit consensus over payload state plus holonomic constraints reduces payload yaw tracking error by roughly 4x compared to wrench-only consensus, indicating that distributed controllers should coordinate state, not just forces.","The same architecture handles payload mass uncertainty and external pushes in hardware experiments, suggesting it can be used outside simulation for load-carrying legged robots.","Increasing ADMM iterations improves consensus quality, but two iterations already provide sufficient coordination for stable transportation, offering a practical tuning guideline.","Communication delays up to 20 rounds with 50% packet dropouts degrade consensus convergence but still allow stable transportation and recovery."],"fun_headline_variants":["Quadruped teams haul payloads with ADMM-coordinated NMPC","Distributed NMPC cuts solve time 23% for quadruped transport","Safety-critical control for cooperative quadruped payload carrying","ADMM consensus over payload state boosts quadruped coordination"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The safety-critical claim rests on the assumption that HOCBF constraints computed on the reduced-order single-rigid-body model, with states from motion capture and kinematic estimation, guarantee safety of the full 18-DoF robots when tracked by the low-level whole-body controller; the paper itself reports brief apparent safety-boundary violations under MoCap noise and strong pushes.","fun_headline_variants_meta":{"raw":{"variants":["Quadruped teams haul payloads with ADMM-coordinated NMPC","Distributed NMPC cuts solve time 23% for quadruped transport","Safety-critical control for cooperative quadruped payload carrying","ADMM consensus over payload state boosts quadruped coordination"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000663,"raw_usage":{"total_tokens":2871,"prompt_tokens":757,"completion_tokens":2114,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":2042}},"tokens_in":501,"tokens_out":2114,"duration_ms":14482,"temperature":1.0,"reasoning_tokens":2042,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T19:17:08.998135+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the two- or three-agent transportation task using onboard state estimation instead of an external motion-capture system, or with a tighter safety margin than the reported threshold, and record the minimum value of the HOCBF safety functions. If any safety function goes negative for longer than a single control step, or the payload yaw error exceeds the roughly 4x margin reported relative to wrench-only consensus, the central safety-critical claim is contradicted.","supporting_citations":[],"review_version":1}