{"id":"79202936-29d9-4ce8-9dc6-0e64df2435f7","arxiv_id":"2507.11345","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"RAE+UPOM, which uses shared hierarchical operational models for both acting and planning, was deployed on a real mobile manipulator and completed object-collection trials under execution failures.","lead":"This paper reports the first physical deployment of RAE+UPOM, a system that lets a reactive acting engine and a Monte Carlo planner share the same hierarchical operational models, on a mobile manipulator. The robot collected objects from tables and completed most trials despite navigation, perception, and arm failures, and the authors analyze how the planner chose actions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Perception-failure adaptation is contradicted by Trials 2.1/3.3: the simulator's frustum check cannot model DOPE confidence failures, so UPOM is overconfident about perception.","rationale":"The reader's weakest assumption identifies the hand-crafted simulation models as the central fragility. I agree, but the paper provides a concrete, self-confessed instance that is more specific and more damaging: the perception simulation ignores the DOPE confidence threshold, causing a whole class of perception failures to be invisible to the planner. This directly contradicts the abstract's 'robust task execution under ... sensor noise' and the reader's paraphrase 'adapting to ... perception failures.' The reader's verdict of CONDITIONAL remains appropriate—the paper does demonstrate a real hardware deployment and successful interleaving, and the concern is addressable by improving the perception model. No change to the verdict is needed, but the conditionality should explicitly include the requirement that the planner's perception model must either model detection failures or the robustness claims must be narrowed. I set verdict_should_be to UNCHANGED because the reader's conditional assessment already captures the need for more evidence; my concern reinforces that condition rather than overturning the overall judgment.","tokens_in":8837,"tokens_out":6144,"duration_ms":74531,"concrete_test":"Instrument the deployed system and the simulator to log, for each perceive(r, o) call, the simulator's predicted outcome (frustum check pass/fail) and the real DOPE confidence and detection success. Then run a trial with the mustard object placed in a pose known to yield DOPE confidence just below 0.5, across the full discrete pan/lift parameter sets. If any camera pose is predicted as successful by the frustum check but yields real DOPE confidence <0.5—or if UPOM ever selects such a pose and the object is missed—then the planner's perception model is overconfident and directly explains the uncollected objects in Trials 2.1 and 3.3. A complementary check is to add a probabilistic failure model for DOPE confidence to the simulator and re-run the same trials; if the planner then avoids those poses, the original failure is attributable to the simulation gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that RAE+UPOM adapts to perception failures—is directly undercut by the paper's own results. Trial 2.1 and Trial 3.3 (Section V) both report objects left uncollected because 'the robot failed to perceive the objects' or 'detection confidence drops below a threshold of 0.5, the object is not recognized at all.' The manuscript explicitly says: 'At present, such cases remain unhandled.' This is not a minor edge case: it is the exact perception-failure mode one would expect under sensor noise, and the abstract's 'robust task execution under action failures and sensor noise' is therefore overstated. The root cause lies in the planner's simulation model. Section III-A describes the simulated 'read current pose' as a geometric visibility check: it succeeds iff all bounding-box vertices lie inside the camera frustum. Section IV-B states that simulated outcomes are 'hand-crafted' and 'provided by human experts.' This model captures whether an object is geometrically visible, but not whether DOPE's learned pose estimator will return a confidence above the detection threshold. As a result, UPOM's rollouts can assign high expected utility to camera poses (pan, lift parameters) that fail in reality, making the planner systematically overconfident about perception. The system only adapts to failures that are observable as command failures (navigation, arm) and retryable; low-confidence detections produce no failure signal, so RAE never learns the object exists. This is precisely the reader's weakest assumption—hand-crafted simulation realism—manifesting as a concrete, admitted gap. The claim of 'adapting to perception failures' is thus unsupported in the strongest sense: the paper proves the opposite for a realistic failure class.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes the first physical deployment of RAE+UPOM, an integrated actor-planner system that shares hierarchical operational models between acting and planning. The system interleaves the Reactive Acting Engine (RAE) with a UCT-like Monte Carlo planner (UPOM) on the Mobipick mobile manipulator for an object-collection task. The authors model perception, manipulation, and navigation through refinement methods, pair each low-level command with a simulated counterpart for rollouts, and report six single-run trials under nominal conditions, spatial perturbation, and induced failures. The paper claims robust execution under action failures and sensor noise and offers observations about planner rollouts via heatmaps. The main contributions are the deployment, the hierarchical operational model, and an empirical study of the interleaved decision-making.","tokens_in":9203,"tokens_out":7591,"duration_ms":88662,"significance":"If the deployment is taken at face value, the paper demonstrates a useful architectural proof-of-concept: a single hierarchical operational model can drive both a reactive actor and a Monte Carlo planner on a physical robot, and the system can recover from certain navigation and arm failures. This is a positive step toward closing the model gap between planning and execution. The authors are candid about some limitations, and the accompanying website may help reproducibility. However, the empirical evidence is too thin to support the paper's broader claims: trials are not repeated, there is no baseline, and the simulator's perception model demonstrably does not capture the failures that occur. The significance of the work is therefore conditional on substantially strengthened evaluation and more careful claims.","major_comments":[{"comment":"The abstract claims 'robust task execution under action failures and sensor noise,' but two of the six trials (2.1 and 3.3) end with objects uncollected because DOPE detection confidence fell below 0.5 and the robot never became aware of the objects. Section V, Trial 3.3 explicitly states 'At present, such cases remain unhandled.' The claim is therefore overstated: the system is robust to observable command failures (navigation, arm) but not to perception failures, which are a form of sensor noise. The conclusion and abstract should be qualified to say that the system handles selected action failures, with perception-failure robustness left as future work.","section":"Abstract; Section V, Trials 2.1 and 3.3"},{"comment":"The simulator used by UPOM models the read current pose command as a geometric visibility check: it succeeds when all bounding-box vertices lie inside the camera frustum. It does not model the learned DOPE pose estimator's confidence threshold. Because the planner's rollouts assume geometric visibility implies successful detection, UPOM is systematically overconfident about perception. This is not a minor implementation detail: it is the mechanism behind the unhandled failures in Trials 2.1 and 3.3, and it contradicts the paper's claim that the hand-crafted simulation outcomes are 'sufficiently realistic for relative comparison.' A concrete test is to add a probabilistic detection model conditioned on viewing distance and angle to the simulator and check whether the planner changes its selected camera poses in the affected trials.","section":"Section III-A; Section IV-B"},{"comment":"Each trial condition was executed once. There are no repeated runs, no variance measures, and no statistical tests, yet the paper draws quantitative comparisons, e.g., 'Using the box enabled the robot to complete the task ∼49 seconds faster' (Section V, Study 1). Single-run results cannot support such comparative claims, and the absence of any baseline or ablation means the contribution of UPOM's planning, as opposed to the hand-coded methods, is not demonstrated. At minimum, multiple runs per condition and a comparison against RAE without UPOM (or random method selection) are needed before claiming empirical insights.","section":"Section V, Table II"},{"comment":"The simulation model is hand-crafted: Section IV-B says costs and outcome probabilities are 'provided by human experts,' and Section V says collection probabilities were 'empirically determined' for the specific object set in the test scenario. Because UPOM's choices are evaluated using this same model, the favorable behavior in the trials may be due to the model being tuned to the scenario rather than to the planner's own merits. This circularity risk is load-bearing for the planner's decision quality. The paper should include a sensitivity analysis over the model parameters, or test a held-out object or table configuration not used in calibration, to show that the planner's choices transfer.","section":"Section IV-B; Section V"}],"minor_comments":[{"comment":"The sentence 'The results aligned with the expectations' is immediately followed by the statement that the robot failed to perceive objects on Table 3 and collected only 2/4 objects. This is contradictory as written and should be rephrased to distinguish the successful adaptation to spacing from the perception failure.","section":"Section V, Study 2"},{"comment":"The paper states in the introduction that the robot 'operates in a fully autonomous manner,' but Trial 3.1 required operator intervention to recover from cable entanglement. The role of operator intervention should be described explicitly in the definition of 'fully autonomous.'","section":"Section V, Trial 3.1"},{"comment":"The values of the decay constant k and the scaling constants c1 and c2 are not reported, and no sensitivity analysis is given. Since c1+c2=1 leaves one free scale, the reader cannot assess how sensitive the planner's choices are to this function.","section":"Equation (1)"},{"comment":"The discrete ranges for SHOULDER_PAN_JOINT and ELBOW_LIFT_JOINT and the value of d_MAX are not specified, so the size of the method-instance space is not reproducible. Please provide these ranges or a reference to the simulator configuration.","section":"Section III-A"},{"comment":"The heatmap description mentions rollout containers, red dots, and bar height, but a reader of the text alone cannot map colors and heights to axes or understand how 'rollout cluster' is defined. Please add axis labels and a legend, or defer details to the caption.","section":"Fig. 5"},{"comment":"The text reports '32.1' for trial 1.2 while Table II lists '32.09'; please make the numbers consistent.","section":"Section V, Study 1"}],"recommendation":"major_revision","confidential_remarks":"The main unresolved question is whether this is indeed the first physical deployment of such an integrated actor-planner system; the 'to our knowledge' claim is hard to verify from the cited literature and might be checked against OMPAS-related hardware work. More importantly, the single-run evaluation is thin for an archival venue. A revised version with repeated trials, a baseline comparison, and a validation of the simulation model's perception component would make the empirical contribution substantially stronger."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful thing here is the concrete integration work: RAE+UPOM, previously only in toy domains and simulation, actually runs on the Mobipick and completes object collection in six trials. The authors also added two sensible engineering updates (utility on partial success, retry counts) and a heatmap visualization of planner rollouts that gives a real look inside the anytime UCT search. That is genuine empirical content, and the paper is honest about the architecture and its limitations.\n\nThe soft spots are real but not fatal. The evaluation is thin: one trial per condition, no error bars, no baseline comparison, and the simulation parameters are hand-populated by human experts. The reader's circularity concern is fair—the planner picks methods based on rollouts from a model tuned for this exact object set—but the real-robot counts do provide some external grounding. The bigger issue, and the one that should bother you most, is the perception gap. The simulator's 'read current pose' is a geometric frustum check; it does not model DOPE confidence falling below threshold. Trials 2.1 and 3.3 both show objects left uncollected because the robot never perceives them, and the paper admits 'such cases remain unhandled.' That is not a minor edge case for an abstract that promises robust execution under sensor noise. The planner is overconfident about perception because the simulation cannot represent low-confidence detections, so the claimed adaptation to perception failures is unsupported and in fact contradicted for this failure class. The authors should either handle that failure mode or explicitly scope the claim to action failures.\n\nEven with that, the paper deserves a serious referee. It is a reproducible system paper with honest reporting, and the underlying shared-model idea is worth debating. For a revision, I would ask for more trials or at least repeated runs, a simple baseline (e.g., greedy or fixed-order collection), and an explicit discussion of where the hand-crafted model diverges from real perception. The citation pattern looks fine—the prior RAE+UPOM work is cited, and the new content is clearly the deployment.\n\nWho is this for? Researchers working on integrated acting and planning, especially in the ROS/refinement-actor community. It is not a breakthrough theory paper, but it is a legitimate engineering study that exposes real mismatches between simulation and reality. I would bring it to the reading group and, with the stated caveats, would accept it for peer review.","headline":"A solid but narrowly evidenced systems paper: first real-robot RAE+UPOM deployment works for action failures, but the abstract oversells sensor-noise robustness and the perception-failure gap needs explicit handling.","tokens_in":9757,"tokens_out":1210,"would_cite":true,"duration_ms":16969,"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":"RAE+UPOM is the first integrated actor-planner system sharing hierarchical operational models to be deployed on a physical robot, and it completes object collection under real failures.","keywords":["hierarchical operational models","refinement acting","Monte Carlo planning","RAE+UPOM","mobile manipulation","object collection","planning-execution integration","action failure recovery"],"falsifier":"Run repeated object-collection trials under an identical layout while setting one object's simulated collection probability much higher than its measured physical success rate; if the planner consistently selects that object first and the realized total utility is lower than an alternative order, the rollouts' relative comparisons are not reliable enough to support the deployment claim.","tokens_in":1608,"feed_emoji":"🤖","tokens_out":1623,"duration_ms":94127,"temperature":0.7,"pith_summary":"The paper is trying to establish that a robot can use one and the same hierarchical operational model for both acting and planning, and that the integrated system RAE+UPOM works on real hardware, not only in toy simulations. It presents the first physical deployment on the Mobipick mobile manipulator, where the robot collects objects from tables despite navigation, perception, and arm failures. If the claim is right, roboticists no longer need to maintain a separate abstract planner model alongside the concrete control code, because the planner can simulate the actual code the actor will run. The experiments also provide a look inside the planner's rollouts, showing how method choices and the emergent use of a transport box arise from expected-utility calculations.","feed_headline":"Real robot plans by simulating the code it will execute","feed_subtitle":"The same operational model drives planning and acting, so real navigation and perception failures are recoverable.","key_machinery":"The central object is a hierarchical operational model: a set of refinement methods, which are computer programs giving alternative procedures for accomplishing tasks, with parameters that can be instantiated in many ways. RAE executes the body of a chosen method instance step by step, while UPOM runs UCT-like Monte Carlo rollouts using a lightweight simulated counterpart of every low-level robot routine, sampling probabilistic outcomes and cost estimates provided by human experts. The coupling that carries the argument is that the actor and planner share the exact same method code, so planning is a simulated execution of the code that will later run on the robot, and the utility function uses an exponentially decaying reward to favor early collection of high-value objects.","core_discovery":"The paper claims that RAE+UPOM, a system that combines the Reactive Acting Engine with the anytime UCT-like Monte Carlo planner UPOM, can be deployed on a physical robot and can complete a hierarchical object collection task under partial observability, sensor noise, and action failures. The actor and planner share the same set of refinement methods, which are Python-coded procedures for tasks such as collecting objects from a table, perceiving an object, and moving objects with or without a box. During execution, UPOM performs Monte Carlo rollouts over alternative method instances using simulated counterparts of the robot's low-level commands, then suggests the method instance with the highest expected utility. The paper reports successful trials with all objects collected, as well as trials where cable entanglement, repeated navigation failure, or missed perception caused partial collection, and it argues that these results demonstrate that planning by simulated execution of operational code closes the gap between descriptive planner models and the real control structures on the robot.","pith_inferences":["If the shared-model approach generalizes, a practical design rule follows: every low-level robot routine should ship with a lightweight simulated counterpart and a success detector, so planning and acting cannot drift apart; the paper flags this as an upfront effort cost, which suggests it could become a standard engineering practice.","Because exploration and perception tasks receive zero reward in the rollouts, the planner is blind to information value, and early planner calls often produce no utility; adding an explicit information-gain term to the utility function could reduce this waste and is a testable extension the paper does not explore.","The two modifications made for real hardware, partial-utility accounting after command failure and bounded retries of the same method, are likely portable to other refinement acting engines even without UPOM, and could be evaluated in simulated factory or fleet tasks."],"forward_implications":["Robots can plan by simulating the same Python-coded refinement methods they will execute, removing the need for a separate PDDL-style descriptive model.","Online planning is fast enough for real interleaving: UPOM used about one second for 100 rollouts per subtask, allowing decisions during execution rather than only before it.","Failures need not abort the mission: with configurable retry counts, the robot returned to a table after three navigation failures and recovered from an emergency stop caused by cable entanglement.","Optional equipment such as a zero-utility transport box is adopted spontaneously when planning discovers that grouping objects raises expected utility.","Simulated camera-frustum checks can select pan and lift joint angles before grasping, and those choices transfer to real perception through a learned 6D pose estimator."],"supporting_citations":[{"why":"Supplies the RAE+UPOM algorithm, the utility model, and the convergence guarantees this physical deployment adapts and extends.","marker":"[2]"},{"why":"Supplies the acting-and-planning principles and the refinement-method formulation on which the shared hierarchical operational models are based.","marker":"[3]"},{"why":"Describes the Mobipick platform whose mobile base, UR5 arm, and camera are used in the real-world trials.","marker":"[4]"},{"why":"Defines the descriptive-versus-operational model gap that motivates the integrated design.","marker":"[1]"},{"why":"Provides a contrasting robot planning architecture that keeps planning and execution knowledge separate, against which the shared-model design is positioned.","marker":"[5]"},{"why":"Describes a comparable refinement acting engine with continuous planning evaluated only in simulation, supporting the first-physical-deployment claim.","marker":"[12]"},{"why":"Supplies the deep-learning 6D pose estimator used at execution time whenever the simulated visibility check predicts that an object can be perceived.","marker":"[13]"}],"fun_headline_variants":["Planning by simulating the robot's own code, on a real robot","Robot executes and plans with the same operational model","Shared hierarchical model enables planning by simulation on a real robot","Monte Carlo planner simulates real robot code to recover from failures","RAE+UPOM: real robot plans by executing simulated code"],"cache_read_input_tokens":11776,"weakest_assumption_plain":"The central load-bearing premise is that the human-supplied simulated models of robot actions are realistic enough for relative comparison, so the method instance with the highest simulated expected utility is also the best choice on the physical robot.","fun_headline_variants_meta":{"raw":{"variants":["Planning by simulating the robot's own code, on a real robot","Robot executes and plans with the same operational model","Shared hierarchical model enables planning by simulation on a real robot","Monte Carlo planner simulates real robot code to recover from failures","RAE+UPOM: real robot plans by executing simulated code"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000633,"raw_usage":{"total_tokens":2869,"prompt_tokens":840,"completion_tokens":2029,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":1944}},"tokens_in":456,"tokens_out":2029,"duration_ms":14356,"temperature":1.0,"reasoning_tokens":1944,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:10:09.113780+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run repeated object-collection trials under an identical layout while setting one object's simulated collection probability much higher than its measured physical success rate; if the planner consistently selects that object first and the realized total utility is lower than an alternative order, the rollouts' relative comparisons are not reliable enough to support the deployment claim.","supporting_citations":[{"cited_title":"Deliberative acting, planning and learning with hierarchical operational models,","cited_arxiv_id":null,"evidence_quote":"Supplies the RAE+UPOM algorithm, the utility model, and the convergence guarantees this physical deployment adapts and extends."},{"cited_title":"Ghallab, D","cited_arxiv_id":null,"evidence_quote":"Supplies the acting-and-planning principles and the refinement-method formulation on which the shared hierarchical operational models are based."},{"cited_title":"A physics-based simulated robotics testbed for planning and acting research,","cited_arxiv_id":null,"evidence_quote":"Describes the Mobipick platform whose mobile base, UR5 arm, and camera are used in the real-world trials."},{"cited_title":"Deliberation for autonomous robots: A survey,","cited_arxiv_id":null,"evidence_quote":"Defines the descriptive-versus-operational model gap that motivates the integrated design."},{"cited_title":"ROSPlan: Planning in the robot operating system,","cited_arxiv_id":null,"evidence_quote":"Provides a contrasting robot planning architecture that keeps planning and execution knowledge separate, against which the shared-model design is positioned."},{"cited_title":"Enhancing operational de- liberation in a refinement acting engine with continuous planning,","cited_arxiv_id":null,"evidence_quote":"Describes a comparable refinement acting engine with continuous planning evaluated only in simulation, supporting the first-physical-deployment claim."}],"review_version":1}