{"id":"65dc084d-befd-44e1-b332-f3334bd720b3","arxiv_id":"2502.02858","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"p-SSA resolves infeasible multi-constraint safe-control QPs by first minimizing constraint violations and then tracking the nominal control, enabling dexterous humanoid collision avoidance in clutter.","lead":"A new safe-control algorithm, p-SSA, keeps a humanoid robot from hitting obstacles or itself by relaxing conflicting safety constraints just enough to keep the control problem solvable. It was tested in simulation and on a real Unitree G1 humanoid, where it reduced collisions during teleoperation without per-task tuning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'minimal safety violations' claim is only about slack in linearized control constraints, not about distance-based safety; Section VI even concedes no safety guarantee, so the central claim needs an explicit link or a softened wording.","rationale":"The two-stage construction in Section IV is mathematically clean: Phase I is always feasible because s can absorb any constraint violation, and Phase II is feasible by construction with the minimal s*. The paper also deserves credit for the explicit infeasibility taxonomy and the breadth of the simulation study. My concern is not with the algorithm's feasibility guarantee but with the semantic claim that this minimal slack corresponds to 'minimal safety violations.' The slack lives in velocity space, and the C metric is defined directly from the solver's slack, so the empirical comparison of C is partially circular; S is the metric that matters, and it is reported without statistical support. Section VI explicitly concedes that no safety guarantee survives relaxation, which is honest but undercuts the wording of the abstract. This is an overclaim, not a fatal flaw: the algorithm may still perform well, and the paper could be repaired by (a) stating the surrogate nature of the Phase I objective, (b) providing S-based comparisons with error bars, and (c) adding the oracle experiment. The velocity-tracking assumption is standard and acknowledged, but the slack-to-safety link is not even acknowledged as an assumption. Therefore I recommend keeping the CONDITIONAL verdict, with revision requirements targeting the claim and evaluation rather than the algorithm itself.","tokens_in":15953,"tokens_out":7948,"duration_ms":81935,"concrete_test":"Re-run the eight simulation tasks with an oracle Phase I that replaces (12)'s objective with the true one-step-ahead safety cost sum_i max(0, dmin - d_i(x + u * dt)) (using the same first-order model and constraint set), then compare the S scores and collision counts of p-SSA against this oracle. If p-SSA's S scores are not statistically indistinguishable from the oracle's (e.g., within noise over the 2000-step runs), the claim that minimizing slack in (12) minimizes actual safety violations is refuted; if they match, the surrogate concern is empirically resolved. Also compute the same comparison with a first-order lag on joint velocities to test the accurate-tracking assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that p-SSA 'always operates within the maximal feasible region' and achieves 'minimal safety violations' rests on an unproven equivalence: minimizing the slack vector s in (12) minimizes the true distance-based safety violations measured by S in (16). This equivalence does not hold in general. For the first-order model, (12b) is -J_i(x)u <= -eta + s_i, i.e., J_i u >= eta - s_i, so s_i is a shortfall in the required separating velocity, not a distance. Integrating that velocity shortfall over time is what produces actual collisions, and the mapping from instantaneous slack to collision depth or duration depends on geometry and discretization; minimizing the p-norm of instantaneous slack can spread violations across many constraints (with p=2) rather than preventing the deepest contact. The C metric in (15) is computed from the very slack variables that p-SSA minimizes, so the reported C advantage is partly circular; S is the honest measure, and the paper gives only point estimates with no error bars or code. Section VI admits that 'as long as relaxation is needed, the safety violation cannot be bounded, preventing any safety guarantee to hold.' Thus the abstract's 'minimizing safety violations' is a claim about a surrogate objective, not about the safety specification. This is a correctness risk in the central claim, distinct from the velocity-tracking assumption: even with a perfect tracker, a relaxed constraint only controls the instantaneous derivative of the safety index, so the chosen u can still produce collisions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses dexterous safe control for humanoid robots in cluttered environments, where a multi-constraint safe control QP with hundreds of collision-avoidance constraints frequently becomes infeasible. The authors propose r-SSA, which adds weighted slack variables to relax the safety constraints, and p-SSA, a two-phase method that first computes a minimal slack vector to render the constraints feasible and then solves a tracking QP with that slack fixed, guaranteeing feasibility of the second QP by construction. The paper reports simulation results on eight tasks with a Unitree G1 model and a real hardware teleoperation demonstration, claiming minimal safety violations, zero parameter tuning, and top performance relative to baselines.","tokens_in":16300,"tokens_out":8435,"duration_ms":79758,"significance":"The Phase I / Phase II decoupling is a clean idea with practical value: it avoids the safety-performance weight tuning of r-SSA and guarantees feasibility of the tracking QP whenever the projection step is solved, which is nontrivial for high-dimensional multi-constraint problems. The real-robot demonstration is also a useful step toward deployable dexterous safety. However, the headline claims are not fully supported. The 'minimal safety violations' assertion concerns a slack variable in a linearized derivative constraint, not the distance-based safety specification, and the reported advantage is partly built into the evaluation metric. The velocity-tracking assumption underlying the hardware experiments is not validated. With a corrected evaluation and a more careful statement of what is minimized, the contribution would be solid.","major_comments":[{"comment":"The 'minimal safety violations' claim is not supported by the evidence as presented. In Eq. (15), the C score for p-SSA uses s_i,t from the Phase I solution of (12), but the executed control is the Phase II solution of (13); the realized control-constraint slack of that control, max(dot-phi_i(x_t,u_t)+eta_i,0), is generally different from the Phase I slack, so the reported C advantage is partly an artifact of the projection rather than of the actual behavior. More importantly, the Phase I objective is a norm of the slack in the linearized derivative constraint (12b), not a measure of distance-based safety violation; no theorem or experiment connects this slack to the S metric in (16). Section VI explicitly concedes that once relaxation is active, no safety guarantee holds. The abstract and Section IV-B should either rephrase the claim to 'minimal violation of the linearized safety constraints' or provide a concrete link, such as a bound S <= F(||s||) for the first-order model.","section":"Section V-D, Eq. (15); Section IV-B, Eq. (12)"},{"comment":"The safety constraints are derived from the first-order model x_dot = u, with the text stating 'assuming an accurate velocity tracker at the downstream.' In the real Unitree G1 experiment, no tracking-error measurements or robustness margins are reported. If the low-level velocity tracker is imperfect, the inequality (9b) does not describe the actual evolution of the safety index, and a feasible p-SSA solution can still permit collisions. The paper should either quantify the tracking error on the real robot, add a margin to dmin to absorb it, or explicitly limit the hardware claim to an empirical demonstration rather than a safety guarantee.","section":"Section V-A and V-G"},{"comment":"The claim of being 'totally tuning-free' / 'zero parameter tuning' is overstated. p-SSA still requires choosing eta, the safety margins dmin,env and dmin,self, the norm order p, and Qpssa_s (set to identity in the paper). What is tuning-free is the relative weight between the tracking objective and the slack penalty, compared with r-SSA's Qrssa_s. Please rephrase to 'without tuning the safety-performance trade-off weight'.","section":"Section IV-B and abstract"},{"comment":"The Phase I optimization is written as minimizing over s only, while u appears in the constraints (12b)-(12c). As written, u is not a decision variable and the problem is not well-posed. To guarantee the Phase II feasibility argument, Phase I must be understood as an optimization over (u,s), or at least as an existential projection over u. Please correct the notation and specify the metric with respect to which the constraint set is projected.","section":"Section IV-B, Eq. (12)"},{"comment":"The headline comparisons are based on a single 2000-step trajectory per task, with no multiple seeds, error bars, or statistical significance tests. Since the dynamic-obstacle tasks use Brownian motion, run-to-run variance is non-negligible, and the Pareto-front plots in Figures 7 and 8 represent point estimates. The claim that p-SSA 'automatically secures the sweet spot' on the Pareto fronts needs at least a few seeds per configuration or a statistical comparison.","section":"Section V-E and V-F"}],"minor_comments":[{"comment":"Typo: 'sprase' should be 'sparse'.","section":"Abstract"},{"comment":"Typo: 'simluated' should be 'simulated' and 'teleopration' should be 'teleoperation'.","section":"Figure 9 caption"},{"comment":"The statement 'the naive SSA essentially ignores the obstacles if the QP is infeasible' is imprecise; SSA actually passes uref unmodified when the QP is infeasible, which may or may not avoid obstacles. Consider rewording.","section":"Section V-E-a"},{"comment":"The C and S metrics are undefined (0/0) when there are no infeasible steps or no distance violations, respectively. Please state the convention used in the plots for these cases.","section":"Section V-D, Eq. (15)"},{"comment":"The caption states that p-SSA 'consistently bounds phi below zero,' but the text acknowledges a tolerance due to discretization. Please quantify this tolerance and the conditions under which it is bounded.","section":"Section IX-C, Figure 10"}],"recommendation":"major_revision","confidential_remarks":"The core feasibility-by-construction idea is sound and likely of interest to the robotics audience. The main risk is that the paper's central evaluative claim is partly circular (C metric computed from the very slack minimized by p-SSA) and, for the S metric, only point estimates with no variance are reported. If the authors can provide a realized-slack version of C, a direct or empirical link from slack to distance-based violations, and multi-seed results, the paper would be much stronger. I would also ask the authors to release code or detailed simulation seeds, since no code is mentioned in the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"p-SSA is a clean, practically motivated two-stage QP formulation that deserves serious referee time, but the \"minimal safety violations\" claim outruns what is actually proven. Phase I minimizes slack in linearized control constraints, not true distance-based safety violations, and the paper's own Section VI concedes that no safety guarantee survives relaxation.\n\nThe dexterous-safety problem setting is real: limb-level geometry, dense obstacles, and hundreds of simultaneous constraints produce infeasible QPs, and the taxonomy of inherent/method/kinematics infeasibility is useful. The two-stage projection idea is elegant: solve for minimal slack first, then solve the tracking QP with that slack fixed. Phase II feasibility by construction is correct, and the simulation effort is substantial (eight tasks, up to 950 constraints, plus a hardware demo on a Unitree G1). The ablation showing that r-SSA needs task-specific tuning while p-SSA lands on a Pareto sweet spot is a decent empirical point.\n\nThe main soft spot is the surrogate-objective issue. Phase I minimizes the p-norm of s in Eq. (12), where s is a shortfall in required separating velocity, not a distance. Minimizing that does not in general minimize collision depth or duration; with p=2 it can spread violations across constraints rather than prevent the deepest contact. The C metric in Eq. (15) is computed from the same slack variables p-SSA minimizes, so the reported C advantage is partly definitional. The S metric is the honest one, and the paper gives only point estimates with no error bars or code. The velocity-tracking assumption (stated in V-A) also means the first-order model x_dot = u may not match the real robot, though that is a standard assumption for kinematic safety filters and not a fatal issue.\n\nThese are fixable. The cleanest remedy is to soften the abstract and Section IV-B wording from \"minimal safety violations\" to \"minimal violation of linearized safety constraints,\" and to report S with error bars or an additional metric based on actual penetration depth. The core idea remains useful: a tuning-free slack-projection filter that keeps the QP feasible is a practical contribution for teleoperation and whole-body control.\n\nI would send this to peer review. The overclaim is real but addressable, and the problem setting plus the simple algorithm give reviewers something concrete to evaluate. For a reader in safe control, it is worth a look; the slack-vs-distance circularity is a good discussion piece.","headline":"Deserves peer review: the two-phase slack-projection idea is clean and the problem setting is real, but the paper overstates a surrogate-objective result as a safety guarantee.","tokens_in":16846,"tokens_out":2732,"would_cite":true,"duration_ms":25646,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C85","90C20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Projected Safe Set Algorithm (p-SSA) splits feasibility projection from task tracking, so a humanoid's multi-constraint safety QP is always feasible and needs no parameter tuning.","keywords":["dexterous safety","safe set algorithm","control barrier functions","humanoid robots","collision avoidance","quadratic programming","infeasible constraints","teleoperation"],"falsifier":"Run p-SSA on a humanoid arm with a deliberately degraded velocity tracker (e.g., 200 ms latency or 30% gain error) inside a narrow cavity, while recording actual collisions; if the QP remains feasible but the robot still hits the obstacle, the accuracy assumption is refuted. Equivalently, construct a two-plane example where the minimal-slack projection moves the end-effector parallel to the planes while the true distance still decreases, showing that minimal constraint slack is not minimal physical violation.","tokens_in":15717,"feed_emoji":"🤖","tokens_out":4884,"duration_ms":41335,"temperature":0.7,"pith_summary":"This paper tackles dexterous safety for humanoids: avoiding external and self-collisions with limb-level geometry in cluttered environments. Such tasks produce hundreds of safety constraints, and the quadratic programs used by standard safe controllers often become infeasible. The authors propose the Projected Safe Set Algorithm (p-SSA), which first solves a slack-minimization phase to project the infeasible constraint set onto the nearest feasible set, then solves the tracking QP with that fixed relaxation. They claim p-SSA always operates within the maximal feasible region, is completely tuning-free, and outperforms the naive SSA and relaxed SSA baselines on all eight simulated tasks, with a real Unitree G1 teleoperation demonstration.","feed_headline":"Two-stage safe filter keeps humanoid control feasible, tuning-free","feed_subtitle":"By minimizing safety violations before tracking, p-SSA handles hundreds of collision constraints on a real Unitree G1.","key_machinery":"The central object is the projection of an infeasible safe-control constraint set onto the nearest feasible set, measured by the $p$-norm of total relaxation. Phase I (12) minimizes that relaxation, choosing slack variables $s^*$; Phase II (13) then solves the original tracking objective $\\|u - u_{\\text{ref}}\\|^2_{2,Q}$ subject to the safety constraints relaxed by exactly $s^*$. Because $s^*$ is feasible for (12), (13) is guaranteed to have a solution. The decoupling removes the trade-off between safety and performance that plagues weighted slack regularization, which is what makes p-SSA parameter-free.","core_discovery":"The paper's central claim is that p-SSA always operates within the maximal feasible region indicated by the given constraint set while being totally tuning-free, and that it achieves top performance across various task settings without parameter tuning. Concretely, the two-stage procedure—minimize the slack $s$ in (12), then solve the tracking QP with the optimal $s^*$ in (13)—guarantees that the second QP is feasible by construction. In experiments p-SSA yields lower control-constraint violation and distance-safety violation scores than naive SSA and r-SSA on all eight simulated dexterous safety tasks, and it prevents collisions in a real Unitree G1 teleoperation test where the operator deliberately makes risky motions.","pith_inferences":["We infer that the slack amount $s^*$ solved in Phase I can serve as an online indicator of how close the current situation is to local infeasibility, something the paper does not exploit.","The paper sets the slack weighting to identity and leaves smart balancing to future work; an obvious test is whether distance- or torque-aware weights push p-SSA's Pareto frontier further.","Because the projection step only needs the constraint set, the same Phase-I/Phase-II split could be applied to CBF-based multi-constraint QPs, which face the same infeasibility problem.","The linearization in the safety constraints means the 'minimal violation' is minimal in constraint space, not necessarily in physical distance; we infer this gap would grow under fast dynamics and imperfect tracking."],"forward_implications":["p-SSA can be deployed directly on high-dimensional humanoid models with hundreds of collision-avoidance constraints without per-task tuning.","When a collision is physically inevitable, p-SSA provides a principled minimal-violation response instead of passing through an arbitrary reference.","The same projection idea extends to any energy-function safe controller that yields a control-affine QP, including CBF-based filters.","Joint-space velocity commands computed by p-SSA can act as a safety filter for teleoperation interfaces, allowing operators to issue unsafe goals while the robot remains safe.","The guarantee that the second QP is feasible holds even under dynamic obstacles, since feasibility comes from the projection, not from the obstacle distribution."],"supporting_citations":[{"why":"Introduces the safe set algorithm the paper extends to multi-constraint settings.","marker":"[15]"},{"why":"Establishes CBF-based QP safety filters that share the same infeasibility issue.","marker":"[2]"},{"why":"Synthesizes safety indices for single-constraint feasibility, the baseline p-SSA avoids needing.","marker":"[30]"},{"why":"Extends safety index synthesis to state-dependent control spaces.","marker":"[5]"},{"why":"Composes multiple control barrier functions but with poor scalability, motivating p-SSA's approach.","marker":"[3]"},{"why":"Unifies energy-function-based safe controllers, supporting the claim that p-SSA generalizes beyond SSA.","marker":"[26]"}],"fun_headline_variants":["p-SSA tames unruly constraints to keep humanoids collision-free","Feasibility-first safe control: p-SSA makes humanoid dexterity possible","Projected Safe Set: handling hundreds of collision constraints with zero tuning","Minimizing safety violations: p-SSA guarantees feasible humanoid control","Tuning-free safety for humanoids: solving infeasible constraint sets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the robot's low-level controller tracks commanded joint velocities accurately enough that the first-order model $\\dot{x}=u$ matches the real robot, so the linearized safety constraints describe the true evolution of the safety index.","fun_headline_variants_meta":{"raw":{"variants":["p-SSA tames unruly constraints to keep humanoids collision-free","Feasibility-first safe control: p-SSA makes humanoid dexterity possible","Projected Safe Set: handling hundreds of collision constraints with zero tuning","Minimizing safety violations: p-SSA guarantees feasible humanoid control","Tuning-free safety for humanoids: solving infeasible constraint sets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000249,"raw_usage":{"total_tokens":1510,"prompt_tokens":864,"completion_tokens":646,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":480,"completion_tokens_details":{"reasoning_tokens":548}},"tokens_in":480,"tokens_out":646,"duration_ms":6303,"temperature":1.0,"reasoning_tokens":548,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T10:50:33.286078+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run p-SSA on a humanoid arm with a deliberately degraded velocity tracker (e.g., 200 ms latency or 30% gain error) inside a narrow cavity, while recording actual collisions; if the QP remains feasible but the robot still hits the obstacle, the accuracy assumption is refuted. Equivalently, construct a two-plane example where the minimal-slack projection moves the end-effector parallel to the planes while the true distance still decreases, showing that minimal constraint slack is not minimal physical violation.","supporting_citations":[{"cited_title":"Control in a safe set: Addressing safety in human-robot interac- tions","cited_arxiv_id":null,"evidence_quote":"Introduces the safe set algorithm the paper extends to multi-constraint settings."},{"cited_title":"Safety index synthesis via sum-of-squares programming","cited_arxiv_id":null,"evidence_quote":"Synthesizes safety indices for single-constraint feasibility, the baseline p-SSA avoids needing."},{"cited_title":"Safety index synthesis with state-dependent control space","cited_arxiv_id":null,"evidence_quote":"Extends safety index synthesis to state-dependent control spaces."},{"cited_title":"Compositions of multiple control barrier functions under input constraints","cited_arxiv_id":null,"evidence_quote":"Composes multiple control barrier functions but with poor scalability, motivating p-SSA's approach."},{"cited_title":"Safe control algorithms using energy functions: A unified framework, bench- mark, and new directions","cited_arxiv_id":null,"evidence_quote":"Unifies energy-function-based safe controllers, supporting the claim that p-SSA generalizes beyond SSA."}],"review_version":1}