{"id":"184fea5d-02df-4ea1-9834-dd3cac6cd3ea","arxiv_id":"2411.10699","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":15,"one_line_summary":"A two-layer controller with adaptive nonlinear MPC and control barrier functions lets teams of quadruped robots safely push objects with unknown mass and friction through cluttered environments.","lead":"This paper presents a hierarchical control system that lets a team of quadruped robots push an unknown object to a target while avoiding static and moving obstacles. It combines an adaptive nonlinear model-predictive planner with safety barriers and per-robot locomotion controllers, tested in simulation and on real Unitree robots.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CBF safety guarantee is not established: constraints are soft penalties and R_i in the robot-obstacle barrier is never defined as a state function, so forward invariance does not follow.","rationale":"The reader's conditional verdict is appropriate: the method is plausible, the simulations and hardware experiments are useful, the code is released, and the engineering contribution stands. However, the strongest reason to withhold unconditional acceptance is not exactly the one emphasized in the reader's weakest_assumption. The missing R_i definition and tracking-error analysis are real, but an even more fundamental issue is that the CBF constraints are converted to soft penalties in Sec. 6.6, so the theoretical forward-invariance guarantee is not realized by the implemented optimization. This is an internal inconsistency with Theorem 1, and it directly undermines the abstract's and conclusion's language of 'guaranteed safety.' The concern is load-bearing because the central claimed contribution is safety; if the safety constraint can be violated by the penalty relaxation, the paper's headline result collapses to 'our planner tends to avoid obstacles in the tested scenarios,' which is weaker than claimed. The suggested test is straightforward because the code is public: log the barrier values along the closed-loop trajectory. If the barrier values remain positive, the penalty relaxation may still be practically sufficient, and the concern would be about the strength of the claim rather than the method's validity. If they go negative, the paper must either enforce the constraints as hard constraints or explicitly reframe the contribution as empirical rather than guaranteed. The Fmax inconsistency in Table 1 (0.7 N vs. 50 N force plots) is a minor additional error that should be fixed in revision but is not central. Overall, the paper is a solid engineering contribution that needs a more honest statement of what is and is not guaranteed.","tokens_in":19647,"tokens_out":3841,"duration_ms":42311,"concrete_test":"In the released open-source repository, locate the OCS2 problem formulation used for the Sec. 7.1.2 simulation and check whether the CBF constraints appear only inside the penalty term l^P (Eq. 50) or as hard constraints in the NLP. Then run that simulation while logging min_{j,i} h^{ri}_{oj}(x_k) and min_j h^m_{oj}(x_k) over the full trajectory, and also record the planned R_i versus the actual robot body position from Gazebo. If any logged h value is negative, or if the solver reports active penalty slack on a CBF term, then the implemented planner does not satisfy the CBF condition and the forward-invariance guarantee of Theorem 1 does not apply.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the hierarchical planner is safety-critical, with CBFs guaranteeing collision avoidance. However, Sec. 6.6 explicitly converts all inequality constraints, including the CBF constraints (26d), into a relaxed log-barrier penalty (47) added to the cost function (48), and the discretized NLP (50) retains only the dynamic equality constraint. Consequently, the implemented planner does not enforce h_cbf(x,u) >= 0 pointwise; it merely penalizes violations. The forward-invariance guarantee of Theorem 1 and Definition 2 applies only when a controller satisfies the CBF inequality, so the theoretical safety guarantee does not hold for the implemented penalty-based formulation. This is an internal inconsistency, not just an implementation detail.\n\nAdditionally, the robot-obstacle barrier (45b) depends on R_i, the robot center position, but R_i is never expressed as a function of the MPC state x = [x_b^T, d^T]^T and input u. Only the contact point r_i is defined in Sec. 2 and Eq. (33). The actual robot body position depends on the decentralized loco-manipulation MPC tracking performance (Sec. 3.3), and no tracking-error bound is provided. The paper therefore does not establish that the safety variable appearing in the CBF corresponds to the true robot position. The hardware and simulation results are encouraging demonstrations, but they do not test the guarantee; they show one successful run and one collision when CBFs are removed. The limitation discussion in Sec. 7.4 acknowledges infeasibility from interaction constraints but does not address the penalty relaxation or the missing R_i definition. These issues should be corrected or the safety claims should be weakened from 'guaranteed' to 'empirically improved'.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hierarchical control architecture for a team of quadruped robots collaboratively manipulating a rigid object with unknown mass, inertia, center-of-mass, and friction parameters. A high-level adaptive nonlinear MPC planner generates manipulation forces and contact-point locations while enforcing stability and safety through CLF and CBF constraints, and a decentralized loco-manipulation MPC tracks the planned contact forces and contact points on each robot. The authors validate the approach in Gazebo simulations with static and dynamic obstacles and in hardware experiments with a Unitree A1 and an Aliengo, and they release the code as open source. The paper argues that the adaptive planner handles unknown object properties and that the CBF layer prevents collisions between the object, the robots, and the environment.","tokens_in":20192,"tokens_out":6196,"duration_ms":63164,"significance":"If the claimed safety guarantees held in the implemented system, this would be a valuable step toward practical multi-robot legged loco-manipulation in uncertain environments. The adaptive parameterization of the object dynamics, the comparative studies with and without the adaptive controller and with and without the robot-level CBFs, and the hardware demonstrations with two different quadruped models are useful contributions, and the open-source code increases reproducibility. However, the central 'safety-critical' claim is not supported by the implementation: the CBF and CLF constraints are converted into soft penalties, and the robot-center quantity appearing in the robot-obstacle barrier is never defined as a function of the planner state. The paper is therefore best read as a strong experimental demonstration rather than a formal safety-guarantee result.","major_comments":[{"comment":"The theoretical safety guarantee does not apply to the implemented planner. The paper states in the introduction and conclusion that the motion planner 'guarantees safety', and Theorem 1 with Definition 2 provide forward invariance when the CBF inequality (9)/(13) is satisfied pointwise. However, the solved NLP (50) retains only the dynamic equality constraint, because all inequality constraints, including h_cbf, are replaced by the penalty cost l_P in (48). Since the discretized problem (49) is not the problem actually solved, the condition required by Theorem 1 is not enforced at the solution, and the safety guarantee is not established. The hardware and simulation ablations (Figs. 7 and 9) show that the penalty term improves behavior in the tested scenarios, but they do not test the guarantee. Please either enforce the CBF constraints in the optimizer, provide a conservative penalty formulation with quantified violation bounds, or revise the 'guaranteed safety' claims to match the implemented soft-constrained formulation.","section":"Sec. 6.6, Eqs. (47), (50), (51)"},{"comment":"The robot-obstacle barrier B^{ri}_{oj} depends on R_i, the robot center position, but the paper never states R_i as a function of the MPC state x = [x_b^T, d^T]^T and input u. The only quantity defined in the state is the contact point r_i in Eq. (33b), while the actual robot body position is produced by the decentralized loco-manipulation controller of Sec. 3.3, whose tracking error is not bounded in the paper. Consequently, the Lie derivative in (46b) is not computable from the planner state, and the safety variable used in the CBF may not correspond to the true robot position. Please provide an explicit expression R_i(q_b, d_i) and a conservative safety margin or a tracking-error bound that ensures the actual robot body remains collision-free when the high-level barrier is satisfied.","section":"Sec. 6.5, Eq. (45b)"},{"comment":"The reported maximum manipulation force is Fmax = 0.7 N, but the force plots in Fig. 5 show forces of roughly 10 to 50 N in the successful simulation run. If the force bound (35) is only a soft penalty through (47)-(48), the planner can violate it at nonzero cost; if Fmax is intended to be 50 N, the table is incorrect. Either way, the parameter set reported in Table 1 is inconsistent with the experimental data, and this inconsistency should be corrected and the penalty parameters (rho, epsilon)_bound explained for the force limit.","section":"Table 1 and Fig. 5"}],"minor_comments":[{"comment":"The presentation is inconsistent: problem (49) lists the inequality constraints h_bound, h_clf, and h_cbf as explicit constraints, but Sec. 6.6 and problem (50) state that they are replaced by penalties. The reader should be told clearly that (49) is the conceptual problem and (50) is the solved problem.","section":"Sec. 6.6, Eqs. (49)-(50)"},{"comment":"The procedure that converts the estimated travel time t_avg and the horizon T into the subgoal configurations x_ref_b is not described. A definition of how subgoals are generated incrementally would improve reproducibility.","section":"Sec. 5.1"},{"comment":"The block structure of the matrix \\bar{D}_i appears to have misaligned sub-block dimensions; the reader should be able to verify that the matrix is R^{15 x 15}.","section":"Eq. (16b)"},{"comment":"The CLF constraint is also implemented as a soft penalty, so the asymptotic stability conclusion of Sec. 6.4 holds only if the constraint is satisfied at every time step; this should be acknowledged alongside the CBF discussion.","section":"Sec. 6.4, Eq. (44)"},{"comment":"The CBF parameters (alpha, beta)_CBF are reported as a single pair, but Eqs. (46a) and (46b) use separate coefficients beta^m, alpha^m, and alpha^{ri} for the object and robot barriers; the table should clarify whether these are identical for all barriers.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the experimental work is solid, but the central safety claim is stronger than the implementation supports. The authors should be asked to either enforce the CBF constraints in the planner or temper the 'guarantee' language, and to correct the Fmax/force-plot inconsistency. The novelty relative to the authors' prior IROS 2023 paper is the safety layer and dynamic-obstacle handling; this is acceptable but should be positioned more precisely in the revised manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid engineering paper with an overclaimed safety guarantee. The contribution is real—an adaptive NMPC planner with CBFs for multi-quadruped collaborative manipulation of unknown payloads, building on the authors' IROS 2023 work by adding obstacle avoidance. They ship code, show comparative simulations with and without the adaptive loop and with/without robot CBFs, validate on hardware with an A1 and Aliengo, and include a dynamic-obstacle sim. That is a useful package and I would not desk-reject it.\n\nThe soft spots are concentrated in the safety argument. Section 6.6 converts all inequality constraints—including the CBF constraints (26d)—into relaxed log-barrier penalties (47), and the discretized NLP (50) retains only the dynamics equality. So the implemented planner does not enforce h_cbf >= 0 pointwise; it penalizes violations. Theorem 1's forward-invariance guarantee therefore does not apply to the thing that is actually solved. This is not a fatal flaw—the penalty approach might be a pragmatic choice—but then the paper should say \"empirically improved safety,\" not \"guarantees safety.\" The limitation section (7.4) even says the interaction constraint is enforced as a hard constraint, which contradicts 6.6; someone needs to reconcile that.\n\nThe second issue: the robot-obstacle barrier (45b) uses R_i, the robot center position, but R_i is never expressed as a function of the MPC state x=[x_b,d] or input u. Only the contact point r_i is defined (Sec. 2, Eq. 33). Whether the actual robot body is safe depends on the low-level loco-manipulation MPC's tracking error, and no bound is given. So the CBF constraint sees something that is not demonstrated to be the robot's true position. Again, the hardware and sim results suggest the approach works, but they don't establish the guarantee.\n\nMinor but real: Table 1 lists Fmax = 0.7 N, while the force plots show values around 25–50 N. A unit or decimal error. Also, the citation pattern is fine; the adaptive control is standard Slotine-Li, and the central planner is distinct from their earlier low-level controller.\n\nBottom line: this is worth a serious referee. The architecture, code, and experiments are valuable to the legged-manipulation community. I'd send it to review and ask for a major revision that either enforces the CBF constraints as hard constraints (at least the robot-obstacle ones) or softens the guaranteed-safety language, defines R_i explicitly, and fixes the Fmax inconsistency. I'd also ask for a short discussion of how much barrier violation the penalty relaxation allows in practice.","headline":"Adaptive NMPC + CBF planner for collaborative quadruped loco-manipulation is a real engineering contribution, but the safety guarantee doesn't follow from the implemented penalty-based constraints.","tokens_in":20586,"tokens_out":2519,"would_cite":true,"duration_ms":25366,"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":"A hierarchical planner with adaptive MPC and control barrier functions lets quadruped teams move unknown heavy objects through static and moving obstacles without knowing mass or friction.","keywords":["safety-critical motion planning","legged robots","collaborative manipulation","loco-manipulation","nonlinear MPC","control barrier functions","adaptive control","obstacle avoidance"],"falsifier":"Run the two-robot hardware obstacle course while motion-capturing both the object and the robot bodies, and compare each robot's measured center position with the planned $R_i$ that feeds the robot–obstacle barrier; if the measured distance to the obstacle ever drops below the barrier radius while the planned barrier stays positive, the planner's safety certificate has been falsified.","tokens_in":19448,"feed_emoji":"🤖","tokens_out":12255,"duration_ms":110391,"temperature":0.7,"pith_summary":"This paper claims that a team of quadruped robots can safely carry a rigid object with unknown mass, inertia, center-of-mass location, and ground friction through an environment containing both static and moving obstacles, without any prior model of the payload. The argument is a hierarchical controller: a high-level nonlinear MPC planner generates collision-free object paths, pushing forces, and contact-point locations, while a decentralized loco-manipulation controller makes each robot track them. Safety is encoded with control barrier functions for both the manipulated object and each individual robot, and an adaptive law with a control Lyapunov constraint compensates for the unknown object properties. Comparative simulations and hardware experiments with a pair of different quadrupeds support the claim, showing that disabling either the adaptive law or the robot-level barrier functions leads to failure or collision. If correct, this moves multi-robot legged manipulation closer to unstructured, human-occupied environments.","feed_headline":"Four-legged robot teams carry unknown loads safely around obstacles","feed_subtitle":"Adaptive planning plus safety barriers lets quadrupeds carry payloads with unknown mass and friction.","key_machinery":"The load-bearing object is an adaptive, safety-critical nonlinear model predictive controller whose decision variables are the pushing-force magnitudes $f_{r,i}$ and contact-point positions $d_i$ for each robot. The planner uses a planar rigid-body model of the object split into a nominal part and a lumped uncertainty term $Y_{\\Psi,r}\\Psi$, estimated online by the update law $\\dot{\\hat{\\Psi}} = -\\Gamma_{\\Psi} Y_{\\Psi,r}^{\\mathsf T} s$, where $s = \\dot{q}_e + \\lambda q_e$ is the composite tracking error. A control Lyapunov constraint built from $V = \\tfrac{1}{2}(s^{\\mathsf T} H s + \\tilde{\\Psi}^{\\mathsf T} \\Gamma_{\\Psi}^{-1} \\tilde{\\Psi})$ keeps the tracking error asymptotically stable, while control barrier functions of relative degree 2 for the object and degree 1 for each robot, of the form $\\|O_j - x_p\\| - R_{j,m} \\ge 0$ and $\\|O_j - R_i\\| - R_{j,r_i} \\ge 0$, force collision avoidance. The hard constraints are relaxed as log-barrier penalties and the discretized problem is solved by sequential quadratic programming.","core_discovery":"On its own terms, the paper establishes that collaborative loco-manipulation by legged robots can be made both adaptive and safety-critical by blending a nonlinear MPC planner with control barrier functions and an adaptive dynamics model. The planner treats the object's configuration and each robot's contact point as states, optimizes perpendicular pushing forces and tangential contact-point velocities, and enforces (i) a CLF constraint that keeps the object tracking error asymptotically stable despite the unknown parameter vector, and (ii) CBF constraints that keep both the object and each robot outside obstacle safety radii. Because the adaptive estimator lumps all model uncertainty into a single regressor–parameter pair, the planner needs no mass, inertia, center-of-mass, or friction values. The paper validates this by comparing planner variants with and without the adaptive controller and with and without robot-level CBFs, and by hardware experiments in which two different quadruped models maneuver a box with an unknown 3 kg load around an obstacle.","pith_inferences":["The safety certificate is only as good as the low-level tracking of the planned contact point; if the robot body lags the planner's $R_i$ by more than the barrier safety radius, the CBF constraint would certify a path that the physical robot does not follow, and the paper does not bound this tracking error.","Because the adaptive law lumps all uncertainty into one parameter vector, simultaneous changes in mass and friction could excite coupling effects; a useful extension would split the parameter vector by physical source and test identifiability.","The planner's CBF-and-CLF machinery is not tied to planar motion, so the same hierarchical design could in principle handle three-dimensional manipulation as long as the interaction remains unilateral and the dynamics remain control-affine.","A natural next experiment is to run the dynamic-obstacle scenario, currently shown only in simulation, on hardware with a person walking through the workspace; the fixed safety radii and replanning rate can then be tested against real motion-capture latency."],"forward_implications":["Teams of quadruped robots could carry payloads through warehouses and construction sites without re-tuning for each object's mass, inertia, center of mass, or ground friction, because those are lumped into the adaptive parameter vector.","The planner must track robot–obstacle safety separately from object–obstacle safety; the hardware comparison shows that removing only the robot barriers lets a robot collide even while the object path remains safe.","Formations with different numbers and models of robots are supported, since the low-level controller is decentralized and the planner only assumes unilateral pushing contacts at measured contact points.","Sharp turns and fast rotations are currently unreliable: because the robots cannot apply torque directly and only push, the hard interaction constraint can make the MPC infeasible for aggressive commands."],"supporting_citations":[{"why":"Gives the control barrier function definition and the forward-invariance theorem that turns collision avoidance into a constraint on the derivative of the barrier.","marker":"Ames et al. (2017)"},{"why":"Provides the exponential control barrier function (ECBF) construction used for the relative-degree-2 object–obstacle barrier.","marker":"Ames et al. (2019)"},{"why":"Supplies the composite-error regressor adaptive control machinery, the skew-symmetry argument, and Barbalat's lemma behind the CLF constraint.","marker":"Slotine and Li (1991)"},{"why":"Provides the decentralized unified loco-manipulation MPC that the simulation and hardware results use as the low-level tracking controller.","marker":"Sombolestan and Nguyen (2023b)"},{"why":"Earlier version of this collaborative manipulation system without safety constraints; this paper extends it with CBF obstacle avoidance.","marker":"Sombolestan and Nguyen (2023a)"},{"why":"Establishes the exponential CBF parameter condition (negative real roots) used when choosing the alpha and beta gains in the barrier constraints.","marker":"Nguyen and Sreenath (2016)"},{"why":"Supplies the single-rigid-body state representation for quadruped dynamics on which the low-level locomotion MPC is constructed.","marker":"Di Carlo et al. (2018)"},{"why":"Provides the skew-symmetry property of the inertia and Coriolis matrices used in the Lyapunov stability analysis for adaptive collaborative manipulation.","marker":"Culbertson et al. (2021)"}],"fun_headline_variants":["Quadruped teams adaptively carry unknown loads with safety barriers","Safety-barrier MPC lets legged robot teams handle unknown object loads","Collaborative quadruped teams adapt to unknown payloads for safe movement","Hardware-tested: quadruped teams carry unknown loads safely"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole safety argument depends on the robot actually being where the planner thinks it is; if the low-level controller's tracking error is larger than the collision margin, the planner's safety checks can pass while the real robot hits the obstacle.","fun_headline_variants_meta":{"raw":{"variants":["Quadruped teams adaptively carry unknown loads with safety barriers","Safety-barrier MPC lets legged robot teams handle unknown object loads","Collaborative quadruped teams adapt to unknown payloads for safe movement","Hardware-tested: quadruped teams carry unknown loads safely"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001433,"raw_usage":{"total_tokens":5785,"prompt_tokens":961,"completion_tokens":4824,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":4751}},"tokens_in":577,"tokens_out":4824,"duration_ms":35233,"temperature":1.0,"reasoning_tokens":4751,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:25:05.145779+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the two-robot hardware obstacle course while motion-capturing both the object and the robot bodies, and compare each robot's measured center position with the planned $R_i$ that feeds the robot–obstacle barrier; if the measured distance to the obstacle ever drops below the barrier radius while the planned barrier stays positive, the planner's safety certificate has been falsified.","supporting_citations":[{"cited_title":"Prentice hall Englewood Cliffs, NJ","cited_arxiv_id":null,"evidence_quote":"Supplies the composite-error regressor adaptive control machinery, the skew-symmetry argument, and Barbalat's lemma behind the CLF constraint."},{"cited_title":"In: Proceedings of the American Control Conference, volume 2016-July","cited_arxiv_id":null,"evidence_quote":"Establishes the exponential CBF parameter condition (negative real roots) used when choosing the alpha and beta gains in the barrier constraints."}],"review_version":1}