{"id":"a2ee170b-df7d-4616-b0a1-b4183555f46b","arxiv_id":"2608.07573","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A projection-retraction MPPI variant that enforces equality and inequality constraints inside sampled rollouts and returns commands that satisfy the closed-chain constraint to numerical tolerance.","lead":"This paper introduces a version of MPPI control that keeps two cooperative robot arms on a closed kinematic chain while also respecting joint limits and obstacles. It projects every sampled velocity onto the constraints, then retracts the commanded state back onto the chain manifold to numerical precision, so a strong task cost cannot push the grasp off.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The inequality guarantee relies on Eq. (5) having a usable clearance direction at every state; the paper's own §III-D documents N∇h≈0, where no tangent-space input can restore margin, and no fallback is defined, so Lemma 1's O(Δt²) bound does not cover the regime where it matters.","rationale":"I read the paper as making two distinct claims: (1) Remark 1's equality exactness, which is a conditional statement about retraction termination and is sound as stated; and (2) Lemma 1's finite-step inequality bound, which is the only theoretical support for the safety margins used in experiments. The second claim is the load-bearing one for practical constraint enforcement, and its proof rests on the assumption that the projection QP (5) is feasible at every control cycle with a usable half-space direction. The paper itself provides the counterexample in §III-D: when N∇h vanishes, no input in the equality tangent space can increase clearance, so a moving obstacle will eventually drive the margin negative, regardless of the O(Δt²) bound. The lack of a fallback is therefore not a minor implementation detail but a gap in the central safety argument. I do not see an internal inconsistency in the projection/retraction algebra, the active-set update, or the MPPI weighting that would invalidate the equality exactness claim in the nominal regime. The active-set convergence proof is also missing but is secondary: even with a proof of finite termination, the structural infeasibility direction remains. The reader's weakest-assumption analysis identifies the same issue, and the paper's honest reporting of the failure supports a conditional rather than outright rejection. Hence the existing CONDITIONAL verdict should stand, with the added requirement that the authors either prove feasibility of (5) under explicit kinematic conditions or implement and evaluate the equality-relaxation fallback they mention in the conclusion.","tokens_in":11680,"tokens_out":5675,"duration_ms":62651,"concrete_test":"At the §III-D failure configuration, compute N(q)∇h(q) and solve the linear program max_{u: J_c(q)u=0, ||u||≤ū} J_h(q)u. If the optimum is ≤0 while h̄(q) is below the safety margin, then no equality-respecting velocity can increase clearance; then rerun the scenario with a fallback that relaxes the equality (as the conclusion suggests) and record whether the margin is restored and whether the closed-chain residual remains bounded. This distinguishes a recoverable rate/tuning issue from a structural violation of Lemma 1's premise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"PR-MPPI's central guarantee is conditional on the projection QP (5) being feasible and on the CBF half-space containing an admissible, margin-improving direction at every visited state. Lemma 1 explicitly assumes feasibility of (5) at every cycle, but the paper's own hardware experiment (§III-D) documents exactly the opposite: near the failure configuration, the projected obstacle-gradient N∇h nearly vanishes, so every velocity that satisfies the closed-chain equality J_c u = 0 has, to first order, no component that increases obstacle clearance. In that configuration the QP may still be feasible in the weak sense (all tangent velocities can satisfy J_h u ≥ -γ h̄ when h̄ ≥ 0), but the feasible set contains no retreat direction, so the CBF condition cannot provide forward invariance. A moving obstacle or an aggressive task command can then drive h̄ below zero, and because the algorithm has no specified fallback, the exactness guarantee of Remark 1 continues to hold for the equality while the advertised inequality enforcement fails precisely when the equality is most binding. The paper honestly reports this structural limitation, but does not amend Lemma 1 or the controller with a declared relaxation strategy, so the O(Δt²) margin bound cannot be converted into a safety guarantee without an additional assumption that is neither stated nor verified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Projection-Retraction MPPI (PR-MPPI), a sampling-based MPC that enforces equality and inequality constraints inside the rollout dynamics rather than through soft penalties. At each rollout step, a sampled velocity is projected onto the intersection of the equality-constraint tangent space and the CBF half-spaces that are active within that tangent space, so that inequality corrections never break the equality. After the MPPI update, the averaged command sequence is filtered once more on its own predicted states, and the first commanded configuration is retracted onto the constraint manifold to a numerical tolerance. The paper contributes a command-level equality property (Remark 1), a finite-step inequality error bound (Lemma 1, O(Δt²)), and experiments on a 14-DoF dual-arm Franka system in simulation and on a Unitree H1-2 humanoid in hardware, including ablations and a randomized obstacle-avoidance study.","tokens_in":12010,"tokens_out":4166,"duration_ms":41221,"significance":"If the claims hold, PR-MPPI is a valuable addition to sampling-based MPC for manipulation: it offers a principled, penalty-free way to embed hard equality constraints (closed kinematic chains) together with CBF inequality constraints inside the same projection, with a clean composition that keeps inequality handling from destroying equality satisfaction. The experiments are well designed, include meaningful ablations, and the hardware demonstration addresses a realistic reactive-avoidance scenario. The paper is also honest in explicitly flagging a structural limitation in Section III-D. The main weaknesses are that the inequality guarantee is conditional on an assumption that the paper's own hardware experiment violates, and that the constants appearing in Lemma 1 are not quantified, so the advertised 'safety margin' argument remains qualitative.","major_comments":[{"comment":"Lemma 1's finite-step margin bound (Eq. 16) is derived under the assumption that the projection QP (5) is feasible at every control cycle. Section III-D explicitly documents a configuration where the projected obstacle gradient N∇h nearly vanishes, so no admissible tangent-space input can increase clearance, and the paper states that this is a structural failure rather than a rate or latency artifact. In such a configuration, once the margin becomes negative the QP (5) is infeasible, and the algorithm has no declared fallback; the O(Δt²) bound therefore does not cover the regime where the inequality constraint is most needed. The authors acknowledge the limitation but do not amend Lemma 1 or the controller with a relaxation strategy. The manuscript should either define a fallback (e.g., selective equality relaxation or a minimum-norm infeasibility handling) and extend the bound to that case, or explicitly restrict the safety claim to states where (5) remains feasible.","section":"Section III-D and Lemma 1"},{"comment":"The constants κ_i in Lemma 1 are existential; the paper does not provide a way to compute or bound them in terms of known quantities (second-derivative bounds, Lipschitz constants of J_c and J_h, the tolerance ε_tol, and the input bound ū). Consequently the statement that the O(Δt²) error is 'small enough to be absorbed by the safety margin' (Section III-B) is not quantitative, and the choice of h_safe = 0.05 rad or 0.02 m is heuristic. As written, the lemma is a qualitative characterization rather than a safety certificate, and the forward-invariant-layer interpretation should be flagged as such, or the paper should supply a constructive bound on κ_i together with a margin-selection rule.","section":"Lemma 1 and Section III-B"}],"minor_comments":[{"comment":"The retraction guarantee holds only 'whenever the iteration terminates.' For a real-time controller, the paper should specify a maximum iteration count and define the behavior if the tolerance is not reached within that budget, otherwise the 30 Hz hardware claim is not fully supported.","section":"Remark 1 / Algorithm 1"},{"comment":"The active-set iteration is stated to terminate and yield the minimizer of (5), but no termination argument is given. Since constraints are only added and never removed, a short proof of convergence (or a citation to a standard active-set QP convergence result) would make the projection step fully rigorous.","section":"Section II-B, Eq. (10)-(11)"},{"comment":"The difference between 'Max. bound viol.' and 'Max. margin pen.' is not defined precisely; please clarify whether margin penetration is measured against the shifted guard boundary or the true bound, and make the same distinction in the text of Section III-B.","section":"Table II and Figure 2"},{"comment":"The sentence 'The 0.05 rad safety shift is selected conservatively' is not supported by a computation. Since Lemma 1 does not provide κ_i, this choice is empirical; consider stating explicitly that the margin is a tuned parameter and reporting sensitivity to it.","section":"Section III-B"},{"comment":"The feasibility of the QP (5) is not discussed in the implementation details. The paper should state how infeasibility is detected in the active-set iteration (e.g., maximum iterations or a residual threshold) and what the controller does in that case.","section":"Section II-A / II-B"},{"comment":"The comparison against MC-MPPI and DQ-MPPI is useful, but a comparison against a CBF-QP filter or Shield-MPPI would more directly isolate the benefit of the proposed projection-plus-retraction design for inequality enforcement.","section":"Section III-C"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and makes a solid contribution to constrained sampling-based MPC. The central equality guarantee is honestly framed as a design property, and the experiments are reproducible in spirit with the promised code and videos. The main reservation is that the inequality safety claim is over-stated relative to the theoretical content: Lemma 1 assumes feasibility of (5) at every state, and Section III-D shows a real failure mode that violates that assumption. The authors have reported the limitation candidly, which I credit, but the theory and the controller need to explicitly account for this case—either by a declared relaxation fallback or by a clearly scoped safety statement. I believe these issues are fixable within the scope of a revision, so major_revision is appropriate rather than reject."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, this paper's actual contribution is real: a single velocity-space projector that nests CBF half-space corrections inside the equality tangent space, so inequality enforcement cannot break the closed-chain equality, plus a Gauss-Newton retraction that makes the returned command satisfy the equality to a declared tolerance. That combination is absent from the closest prior work (DQ-MPPI, MC-MPPI, Shield-MPPI), and the paper argues that gap clearly. Second, the paper's central guarantee is honest but conditional, and the authors say so themselves: Remark 1 states the equality exactness is a stopping-rule design property, and Lemma 1's O(Δt²) margin bound explicitly assumes the projection QP (5) is feasible at every cycle with a usable clearance direction.\n\nWhat it does well: the active-set projection in (10)-(11) is sound and solved only over violated or banded margins, which is why compute stays at ~19 ms versus ~239 ms for DQ-MPPI on the same hardware. The ablation design is genuinely informative—execution-only inequality projection still enforces the constraints but degrades search, isolating the value of projecting inside rollouts. The simulator results are clean, and the hardware demo is real: the Unitree H1-2 arms avoid a moving obstacle reactively. The paper also reports the structural failure mode in §III-D plainly: when N∇h nearly vanishes, no tangent-space input can restore clearance, and no fallback is defined. That is a load-bearing limitation, not a hidden one.\n\nSoft spots, in proportion. The feasibility assumption in Lemma 1 is not decorative. The paper's own hardware section documents a configuration where the feasible set contains no retreat direction, so the CBF condition cannot provide forward invariance there. The O(Δt²) bound is not made quantitative: constants κ_i are never computed, and the safety margin is chosen conservatively rather than derived from the bound. The active-set convergence proof is also missing; termination is asserted, not shown. None of this sinks the method—the equality guarantee is a design property and the inequality layer is a practical safety filter—but the claim \"inequality enforcement with exact equality maintenance\" should be read as \"inequality enforcement when a clearance direction exists in the tangent space.\" The authors acknowledge this in the conclusion, though not in the abstract.\n\nThe citation pattern looks fair, and the self-citation to MC-MPPI is not a problem because that baseline is clearly separated in the experiments. The paper is worth a serious referee: it is clearly written, the algorithmic combination is new, and the failure mode is honestly documented. I would send it to review with a request for a declared infeasibility fallback or a stated relaxation strategy, and ideally a quantitative bound on the invariant-layer width. For a robotics audience this is a useful, citable contribution.","headline":"A genuinely new nested projection-retraction MPPI for dual-arm equality plus inequality constraints, with honest reporting of its structural limit; conditional accept with a real soft spot at infeasible clearance directions.","tokens_in":12503,"tokens_out":703,"would_cite":true,"duration_ms":8596,"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":"PR-MPPI enforces hard equality and inequality constraints inside MPPI rollouts, returning commands that satisfy the closed-chain equality to numerical tolerance independent of task weighting.","keywords":["model predictive path integral","MPPI","constraint manifold","control barrier functions","null-space projection","retraction","dual-arm manipulation","closed kinematic chain"],"falsifier":"Run the controller from a state such as the one reported in Section III-D where $N(q)\\nabla \\bar{h}(q)$ is nearly zero and the closed-chain equality forbids the retreating motion; if no feasible projection exists, the controller has no command that satisfies both constraints, contradicting the claim that constraints are enforced to tolerance in every state.","tokens_in":11497,"feed_emoji":"🤖","tokens_out":6329,"duration_ms":55665,"temperature":0.7,"pith_summary":"The paper seeks to show that MPPI, a sampling-based model-predictive controller, can enforce hard constraints by putting them in the sampled dynamics rather than the cost. PR-MPPI projects every sampled velocity into the tangent space of the equality constraint manifold and, within that subspace, into the half-spaces of active inequality constraints, then retracts the returned command onto the manifold. If the construction works, manipulation controllers can keep a two-arm closed chain exact and respect joint limits and obstacle clearances even when an aggressive task objective would overwhelm soft penalties. This matters because MPPI's gradient-free parallel sampling is attractive for nonconvex manipulation costs, but its standard form has no mechanism for constraints that must hold throughout the motion.","feed_headline":"Hard constraints hold to 1e-10 in MPPI rollouts","feed_subtitle":"Projection inside rollouts plus retraction keeps grasp constraints exact, even under aggressive task costs.","key_machinery":"The load-bearing object is the active constraint projector $\\Pi_{\\text{act}}$ of Eq. (5), which solves $\\min_{u} \\tfrac{1}{2}\\|u-\\tilde{u}\\|^2$ subject to $J_c(q)u=0$ and $J_h(q)u\\geq -\\Gamma \\bar{h}(q)$. It is implemented as a null-space projection $N(q)=I-J_c^\\dagger J_c$ followed by an active-set correction that adds violated control-barrier-function half-spaces one at a time, each correction term carrying $N$ as a left factor so the equality stays satisfied to first order. The second mechanism is the retraction map $R_{\\mathcal{M}}$, a Gauss-Newton iteration $q\\leftarrow q-J_c^\\dagger(q)c(q)$ run until $\\|c(q)\\|<\\varepsilon_{\\text{tol}}$, which removes the $O(\\Delta t^2)$ drift of a finite step. These two pieces, projection inside the rollouts and retraction of the executed command, carry the exactness and margin claims.","core_discovery":"The central claim is that a single velocity-space projection, applied at every rollout step and again during execution, lets MPPI maintain both an equality constraint $c(q)=0$ and inequality margins $h(q)\\geq 0$ without turning them into cost penalties. Because each inequality is corrected within the null space of the equality Jacobian, an active margin can never push the state off the equality tangent subspace. The finite-step drift left by this first-order projection is then removed by a Gauss-Newton retraction onto $\\mathcal{M}$, so whenever the retraction terminates the returned command satisfies $\\|c(q^+)\\| < \\varepsilon_{\\text{tol}}$ independently of the task cost, the MPPI temperature and covariance, and the inequality parameters (Remark 1). Lemma 1 bounds the residual inequality error as $O(\\Delta t^2)$, small enough to be absorbed by a fixed safety margin. The paper supports this with a 14-DoF dual-arm simulation study and a humanoid hardware demonstration of reactive obstacle avoidance.","pith_inferences":["Inference: because the projection is purely kinematic, the same nested construction should transfer to sampled torques by using a dynamically consistent null-space projection, moving the exactness guarantee to the actuation layer; the paper leaves this extension to future work.","Inference: near configurations where the projected obstacle-margin gradient nearly vanishes, relaxing the grasp equality with a priority ordering would restore feasibility exactly where the current guarantee is vacuous, at the price of a small commanded grasp slip.","Inference: the $O(\\Delta t^2)$ margin-layer bound suggests a quantitative tuning rule, namely shrinking the safety margin as $\\Delta t$ decreases, which the paper does not state explicitly."],"forward_implications":["The returned joint command satisfies every channel of the closed-chain equality to the commanded tolerance, independent of how strongly the task cost pulls toward a constraint boundary.","Inequality margins are kept to an $O(\\Delta t^2)$ finite-step layer, so a conservatively chosen safety buffer $h_{\\text{safe}}$ is enough to keep measured states outside true limits.","Task costs no longer need soft constraint penalties; in the joint-limit stress test Full PR-MPPI runs without a limit-violation penalty and still never crosses the bound.","Projecting inside the rollouts, not only at execution, makes the optimizer average over feasible samples, which the randomized obstacle task links to a higher completion rate.","The same pipeline runs online at 30 Hz on a humanoid, reactively reshaping both arms around a moving obstacle while the tray remains level."],"supporting_citations":[{"why":"Introduces model predictive path integral control, whose cost-weighted averaging over sampled rollouts PR-MPPI inherits.","marker":"[1]"},{"why":"Provides the information-theoretic MPPI update and parallel-computation formulation used as the base controller.","marker":"[2]"},{"why":"Supplies the control-barrier-function shielding idea for inequality constraints that PR-MPPI adapts into half-space constraints on velocity.","marker":"[5]"},{"why":"Is the closest equality-constrained baseline, projecting samples onto the null space of the closed-chain constraint and compared in both experiments.","marker":"[17]"},{"why":"Is the learned-manifold baseline that PR-MPPI is compared against in the randomized obstacle-avoidance experiment.","marker":"[18]"},{"why":"Provides the MuJoCo physics engine in which the 14-DoF dual-arm simulation experiments are run.","marker":"[19]"}],"fun_headline_variants":["Projection-retraction MPPI enforces constraints exactly","MPPI with exact constraints, no soft penalties","PR-MPPI: project each rollout step, retract the command","Exact constraint-manifold control via projection-retraction","MPPI holds grasp constraints to numerical tolerance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The exactness guarantee assumes that the projection in Eq. (5) is feasible at every state the controller visits; Section III-D documents a hardware configuration where the projected obstacle-margin gradient nearly vanishes, so no admissible velocity can increase clearance without breaking the grasp, and the paper specifies no fallback.","fun_headline_variants_meta":{"raw":{"variants":["Projection-retraction MPPI enforces constraints exactly","MPPI with exact constraints, no soft penalties","PR-MPPI: project each rollout step, retract the command","Exact constraint-manifold control via projection-retraction","MPPI holds grasp constraints to numerical tolerance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000364,"raw_usage":{"total_tokens":1997,"prompt_tokens":1020,"completion_tokens":977,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":898}},"tokens_in":636,"tokens_out":977,"duration_ms":8569,"temperature":1.0,"reasoning_tokens":898,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:32:54.277410+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the controller from a state such as the one reported in Section III-D where $N(q)\\nabla \\bar{h}(q)$ is nearly zero and the closed-chain equality forbids the retreating motion; if no feasible projection exists, the controller has no command that satisfies both constraints, contradicting the claim that constraints are enforced to tolerance in every state.","supporting_citations":[{"cited_title":"Shield model predictive path integral: A computationally efficient robust MPC method using control barrier functions,","cited_arxiv_id":null,"evidence_quote":"Supplies the control-barrier-function shielding idea for inequality constraints that PR-MPPI adapts into half-space constraints on velocity."},{"cited_title":"Real-time dual-arm cooperative manipulation under multiple constraints: A two-stage sampling MPC approach,","cited_arxiv_id":null,"evidence_quote":"Is the closest equality-constrained baseline, projecting samples onto the null space of the closed-chain constraint and compared in both experiments."},{"cited_title":"Manifold-Constrained MPPI: Real-Time Sampling-Based Control Under Hard Constraints","cited_arxiv_id":"2605.24813","evidence_quote":"Is the learned-manifold baseline that PR-MPPI is compared against in the randomized obstacle-avoidance experiment."}],"review_version":1}