{"id":"8dff2d53-65b8-484c-84fc-a8a9cc172200","arxiv_id":"2511.17798","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A priority-ordered whole-body predictive controller that co-solves ORCA-based human-reaction predictions with task plans, validated on two mobile manipulators, beats weighted-sum and open-loop-prediction baselines.","lead":"This paper combines two existing robotics ideas—strict task priorities and predictions of how people react to a robot—into one real-time controller for mobile manipulators, and tests it on two robot platforms with people walking nearby. The result is a single optimization that plans the robot's base-and-arm motion together with each person's likely evasive movement.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Safety claim hinges on humans solving ORCA; no evidence of robustness to non-reciprocal human behavior.","rationale":"The reader's weakest assumption precisely identifies the ORCA-based human model as the load-bearing premise. The paper's mathematical formulation is internally consistent: the KKT conditions (5h)-(5k) provide a necessary and sufficient characterization of the ORCA solution (since the lower-level problem is convex), and the DT-CBF constraint (5f) is a standard discrete-time barrier condition. However, the safety and efficiency claims are only meaningful if the model's predictions match real human behavior. The paper does not report any attempt to validate the ORCA assumption against human participants who are not following reciprocal collision avoidance, nor does it quantify the effect of prediction error on the CBF safety guarantee. The adversarial experiment is presented as a success, but it actually demonstrates that the controller trusts the ORCA prediction even when the human's intention is to block, which is exactly the scenario where the prediction could be wrong. Without either (a) evidence that the DT-CBF constraint remains safe under mispredicted human trajectories, or (b) a clear scoping of the method to cooperative/reciprocal humans, the central claim 'safer' relative to baselines is overstated. This concern is load-bearing because it undermines the main safety contribution, independent of sample-size or baseline-tuning issues. Our recommendation is to keep the CONDITIONAL verdict, adding a requirement to demonstrate robustness to non-ORCA human behavior or to explicitly limit the claims to environments where humans are known to use reciprocal avoidance.","tokens_in":12682,"tokens_out":3344,"duration_ms":35522,"concrete_test":"Re-run the pick-and-place experiments (Sec. V-B) and the adversarial scenario (Sec. V-C) with human participants explicitly instructed to walk along a pre-determined straight-line path (or to ignore the robot entirely), so that their motion is deliberately non-ORCA. Measure the minimum robot-human distance and the number of DT-CBF constraint violations evaluated on the true (measured) human states, comparing against the original ORCA-compliant instructions. If the minimum distance remains above the safety threshold in all trials, the safety claim is robust to prediction error; if not, the claim is only valid under reciprocal human behavior.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central contribution is a controller that couples robot plans with predictions of human motion, where humans are modeled as solving the ORCA problem (Eq. 4). The KKT conditions (5h)-(5k) make the planner co-solve the human's optimization, and the DT-CBF constraint (5f) is evaluated on the resulting predicted human states x^(j). This means the forward-invariance guarantee of h(x) only applies to the predicted human trajectory, not to the true human state. If a real human does not behave according to ORCA (e.g., walks straight without yielding, or actively blocks the robot), the predicted human positions are incorrect, and the CBF constraint can be satisfied while the actual robot-human distance falls below the safety threshold. The paper's only 'adversarial' experiment (Sec. V-C) actually illustrates this vulnerability: although the human is described as deliberately walking toward the robot, SM2ITH plans to continue toward the target because it assumes the human will adapt (Fig. 7b). No documentation of how participants were instructed to behave is provided, so the experimental results may have made the ORCA assumption self-fulfilling. Thus, the advertised 'safe, collision-free' performance is a model-in-the-loop guarantee, not a guarantee against real human behavior. This is more load-bearing than statistical weaknesses in the experiments because even a perfectly powered study would not resolve the mismatch between the model's assumption and true human behavior.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SM2ITH, a bilevel model predictive control framework for mobile manipulators that couples hierarchical task execution (HTMPC) with interactive human motion prediction based on the ORCA model. The robot's lexicographic task optimization is augmented by KKT conditions of the human's ORCA problem, and discrete-time control barrier functions are added to enforce safety with respect to predicted human and obstacle states. The framework is evaluated on two mobile manipulators (Stretch 3 and Ridgeback-UR10) in task-prioritization, pick-and-place, and adversarial human-interaction experiments, totaling 140 runs, with comparisons against weighted-sum MPC, constant-velocity prediction, and reactive-only baselines.","tokens_in":12906,"tokens_out":5332,"duration_ms":54789,"significance":"If the ORCA assumption on human behavior holds, SM2ITH is a plausible and timely integration of hierarchical task MPC and interactive human prediction, extending prior work from mobile navigation to mobile manipulation. The mathematical formulation is clean and built on standard components: the lower-level ORCA problem is convex, so KKT replacement is justified, and the DT-CBF constraint is textbook. The experimental comparison against constant-velocity and reactive baselines is thoughtful, and the two-platform validation adds credibility. However, the paper's central safety claim is not yet substantiated for real human behavior, and the comparative empirical claims rest on small samples without statistical testing. The contribution is incremental relative to the authors' prior works [3] and [11], but the integration itself is non-trivial and of interest to the mobile-manipulation community.","major_comments":[{"comment":"The DT-CBF constraint (5f) is evaluated on x_{k+1}, which includes human states propagated by the ORCA lower-level model (5g)-(5k). The forward-invariance guarantee therefore applies to the model's predicted human states, not to a real human who does not solve ORCA. The abstract's 'safe and collision-free' claim and the adversarial experiment in V-C depend on this assumption: in Fig. 7b, SM2ITH plans to continue toward the target because it predicts the human will yield. The paper does not document how participants were instructed to behave, and no experiment is reported in which the human ignores the robot or blocks non-reciprocally. As written, safety is a model-in-the-loop claim. Please add stress tests with non-ORCA humans (e.g., a scripted straight-line walker, or an active blocker) and/or rephrase all safety claims as conditional on the ORCA assumption, and state participant instru","section":"IV-C, Eq. (5f)-(5k); V-C"},{"comment":"The empirical claims 'SM2ITH performs significantly better' (V-A) and 'balances safety and efficiency' (V-B) are based on n=5 per condition in V-A and n=5-10 per platform/condition in V-B, with heavily overlapping box plots. No significance tests, effect sizes, or confidence intervals are reported. Add appropriate paired statistics (e.g., Wilcoxon signed-rank or permutation tests with multiple-comparison correction) and report the underlying per-run data; otherwise the comparative conclusions are not supported.","section":"V-A/V-B, Figs. 3 and 6"},{"comment":"The weighted-sum baseline uses weights W_ee and W_base described only as 'chosen to approximate hierarchical preferences.' The actual values and any tuning procedure are not given, and no sensitivity analysis is provided. Since a headline result is that strict lexicographic prioritization outperforms weighted-sum MPC, the comparison is interpretable only if the baseline weights are systematically optimized or at least varied over a range. Please report the weights used and show how the conclusions depend on them.","section":"V-A, weighted-sum baseline"}],"minor_comments":[{"comment":"There are broken figure references in the text ('Fig.??' appears three times in Section V-A). Please fix the cross-references.","section":"V-A"},{"comment":"The notation dot q_t in the acceleration constraint is not defined. Clarify whether it is the human's current velocity, the robot's velocity, or a relative quantity.","section":"Eq. (4c)"},{"comment":"The function h is described as a concatenation of constraints but is used in (5f) as if it were scalar. Clarify that (5f) is enforced elementwise, and state the relative degree of h required for the DT-CBF guarantee to apply to this system.","section":"IV-C"},{"comment":"No real-time computation times or solver iteration counts are reported. For an MPC framework advertised for real-time control, please report per-step solve times for both platforms.","section":"V"},{"comment":"The abstract mentions 'adversarial human behavior,' but Section V-C is explicitly qualitative and does not report quantitative metrics. Consider softening the abstract or adding quantitative adversarial metrics.","section":"Abstract / V-C"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically sound as far as the derivation goes, but the safety claim overreaches relative to the ORCA assumption and the experiments are underpowered for the stated comparative conclusions. The stress-test concern about non-ORCA humans is real and should be addressed by reframing claims and/or adding targeted experiments. I do not see a fatal mathematical flaw; the issues are fixable within a revision. Prior-work overlap ([3], [11]) means the novelty is mainly in integration and experiments, which is acceptable for a robotics venue if the claims are calibrated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core formulation is the real contribution: lexicographic task MPC wrapped around a bilevel human-prediction problem, flattened through KKT conditions, with DT-CBF constraints added on top. That assembly is not in the prior literature for mobile manipulation, and the paper carries it through cleanly. ORCA is a convex QP, so the KKT flattening is legitimate, the DT-CBF constraint is textbook, and the lexicographic decomposition inherits the authors' earlier HTMPC work. The math holds up.\n\nCredit is due for the hardware validation: 140 runs total, two platforms with different kinematics, and a video. The banner-versus-cup priority-switching task is a nice way to show that strict hierarchy changes robot behavior in Anthropic terms. The comparison against weighted-sum MPC and against open-loop/reactive human models is the right baseline set.\n\nNow the soft spots, in proportion. The load-bearing one is the safety claim. The CBF forward-invariance guarantee applies to the predicted human trajectory, not the true one. If a human does not solve ORCA — walks straight at the robot, or actively blocks — the predicted states are wrong and the actual distance can fall below threshold. The adversarial experiment in Sec. V-C was supposed to test this, but the paper describes SM2ITH planning toward the target because it assumes the human will adapt (Fig. 7b). That is model-in-the-loop behavior, not robust safety. And without documentation of how participants were instructed, the ORCA assumption may well have been self-fulfilling. This is not a fatal flaw in the formulation — it is a missing caveat in the claim — but it is more serious than the statistical issues, because more data would not fix the mismatch.\n\nThe empirical statistics are soft: n=5–10 per cell, no significance tests, overlapping box plots, and scenario-dependent hand-chosen weights for the weighted-sum baseline. The paper says SM2ITH \"performs significantly better\" at three humans but never tests significance. Those are addressable with code, data, and a few tests. There are also broken figure references (Fig.??) that suggest the manuscript was compiled in haste.\n\nThe citation pattern is honest: the two core building blocks are the authors' own prior work, and they say so. No hidden circularity in the derivation.\n\nWho is this for? Anyone working on task-prioritized whole-body MPC in human environments. It deserves a serious referee — the formulation is valuable and the experiments are ambitious.\n\nRecommendation: send to peer review, but insist the revision report participant instructions, add significance tests or remove the word \"significantly,\" publish code/data, and reframe the safety guarantee as conditional on the ORCA prediction model.","headline":"A clean, useful integration of HTMPC and SICNav for mobile manipulators, with real hardware experiments and sound math; the empirical safety claim is softer than advertised because the CBF guarantee runs on predicted, not real, human states.","tokens_in":13534,"tokens_out":1373,"would_cite":true,"duration_ms":15416,"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 mobile manipulator can keep strict task priorities while treating nearby humans as reactive planners, and the coupled optimization beats weighted-sum and open-loop baselines in tests.","keywords":["mobile manipulation","bilevel model predictive control","interactive human prediction","optimal reciprocal collision avoidance (ORCA)","control barrier functions","task prioritization","human-robot proxemics","lexicographic optimization"],"falsifier":"Record a participant's actual velocity choices while the robot approaches and test whether they solve the embedded ORCA optimization (e.g., by checking KKT residual or comparing to the predicted action). A run where a participant deliberately walks straight into the robot without yielding, while the controller still claims a collision-free guarantee, would break the central assumption; equally, showing that a constant-velocity predictor matches real human behavior as well as the interactive model would undercut the claimed benefit.","tokens_in":12476,"feed_emoji":"🤖","tokens_out":6601,"duration_ms":59669,"temperature":0.7,"pith_summary":"The paper is trying to establish that a mobile manipulator can do two hard things at once: honor a strict priority ordering between navigation and manipulation tasks, and move safely among people who are themselves reacting to the robot. Its answer is to fold a model of human velocity choice—each human solves a small optimal reciprocal collision avoidance problem—into the robot's model predictive controller as lower-level constraints. The combined bilevel problem is made tractable by replacing the human's optimization with its KKT conditions and adding discrete-time control barrier functions for safety. On two different mobile manipulators, across delivery, pick-and-place, and adversarial settings, the method reports tighter tracking of high-priority tasks, shorter detours, and better human-robot proxemics than weighted-sum or open-loop/reactive baselines. If correct, this would let designers encode robot behavior through task priorities—hold the banner versus carry the cup—while still negotiating shared space with humans rather than treating people as moving obstacles.","feed_headline":"Bilevel MPC lets robots treat humans as planners, not obstacles","feed_subtitle":"By assuming humans solve a reciprocal collision-avoidance problem, the robot keeps task priorities in crowded settings.","key_machinery":"The load-bearing object is the bilevel optimization: an upper level that solves the robot's task hierarchy one task at a time, with constraints guaranteeing higher-priority tracking errors do not increase, and a lower level in which each human's velocity is the solution of an ORCA (optimal reciprocal collision avoidance) problem—each agent picks a velocity by minimizing deviation from a desired velocity subject to collision-avoidance half-planes. The lower-level problems are replaced by their KKT conditions (stationarity, complementarity, and dual feasibility), which turns the bilevel problem into a single-level nonlinear program. Discrete-time control barrier functions, one per human and ob","core_discovery":"SM^2ITH's central claim is that a single bilevel MPC can simultaneously enforce strict task priorities and closed-loop human motion prediction. The robot optimizes a lexicographic sequence of task costs (navigation and end-effector) while, at the same time, co-solving the velocity-optimization problems that nearby humans are assumed to be solving—humans are modeled as optimal reciprocal collision avoidance agents—by replacing those lower-level problems with their KKT conditions. Discrete-time control barrier functions on robot-human and robot-obstacle distances provide the safety layer. In delivery, pick-and-place, and adversarial experiments on two mobile manipulators, this combined formula","pith_inferences":["The safety guarantee is model-in-the-loop: the control barrier function is checked against predicted human positions, so the advertised collision-free behavior is only as strong as the ORCA assumption; an explicit bound or estimator for non-reciprocal human behavior would be needed for a formal guarantee against real pedestrians.","The KKT reformulation is what makes the bilevel problem tractable; replacing ORCA with a richer but non-convex human model (social forces, learned predictors) would break this direct embedding and require a different solution scheme.","The task hierarchy itself can act as a communication channel: by choosing whether the arm or the base yields, the robot reveals its intent to humans, which the ORCA model can exploit implicitly through the predicted human velocities.","The cautiousness parameter gamma could be made adaptive—increasing when the human's measured velocity deviates from ORCA-optimality—as a testable extension that might reduce freezing in dense crowds."],"forward_implications":["Strict lexicographic priorities let the same controller produce two qualitatively different avoidance behaviors—move the base to protect a banner, retract the arm to protect a cup—without re-tuning weights.","Closed-loop human prediction yields shorter detours and lower runtime than constant-velocity prediction, and less time inside a human's intimate space than reactive-only control.","In adversarial situations, interactive prediction lets the robot continue toward its target instead of retreating indefinitely, because it anticipates the human will adapt.","Higher human densities (three humans) reveal the largest gap over weighted-sum MPC on the high-priority task, suggesting the hierarchy matters most when conflicts are frequent.","The same formulation transfers to platforms with different kinematics (nonholonomic 7-DOF and holonomic 9-DOF), indicating the approach is platform-agnostic."],"fun_headline_variants":["Bilevel MPC lets robots anticipate human motion for safe tasks","Robot treats humans as reciprocal planners, not obstacles","Bilevel MPC keeps robot priorities while predicting humans","Single optimizer handles robot tasks and human motion prediction","Human-aware MPC: strict priorities and interactive prediction"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Humans walking in the experiments are assumed to actually solve the same ORCA optimization the robot embeds, so their real velocity choices satisfy the KKT conditions; if a person walks non-reciprocally or ignores the robot, the coupled prediction is misspecified and the CBF's collision-free guarantee only holds for the predicted states, not the real ones.","fun_headline_variants_meta":{"raw":{"variants":["Bilevel MPC lets robots anticipate human motion for safe tasks","Robot treats humans as reciprocal planners, not obstacles","Bilevel MPC keeps robot priorities while predicting humans","Single optimizer handles robot tasks and human motion prediction","Human-aware MPC: strict priorities and interactive prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000834,"raw_usage":{"total_tokens":3470,"prompt_tokens":729,"completion_tokens":2741,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":2668}},"tokens_in":473,"tokens_out":2741,"duration_ms":17718,"temperature":1.0,"reasoning_tokens":2668,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T20:54:04.389946+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a participant's actual velocity choices while the robot approaches and test whether they solve the embedded ORCA optimization (e.g., by checking KKT residual or comparing to the predicted action). A run where a participant deliberately walks straight into the robot without yielding, while the controller still claims a collision-free guarantee, would break the central assumption; equally, showing that a constant-velocity predictor matches real human behavior as well as the interactive model would undercut the claimed benefit.","supporting_citations":[],"review_version":1}