{"id":"71e484bb-080a-4f30-93b1-e8f1cac2b60d","arxiv_id":"1908.10227","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A PDDL+-based hybrid task and motion planner computes Kalman-filter belief updates during search and produces uncertainty-aware navigation plans for a mobile robot in a simulated corridor.","lead":"This paper integrates task-level planning with motion planning under uncertainty for robot navigation, using the PDDL+ planning language to simulate belief updates during search. A generalist reader may care because it shows a way to make high-level robot plans aware of positional uncertainty, not just distances and obstacles.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Planner optimizes a single-observation sample path, not the expectation in cost (10), so the claimed minimization is unsupported.","rationale":"The reader's verdict is CONDITIONAL, which remains appropriate. My stress-test identifies a load-bearing concern that is distinct from but complementary to the reader's stated weakest assumption. The reader's weakest_assumption is discretization accuracy (Section IV-D and V), which the authors themselves acknowledge and which can be mitigated by finer discretization. My concern is that the planner's optimization objective is not the expectation in (10) because only one observation sample is used per belief update (Algorithm 1, Section IV-D). This is a more direct threat to the central claim that the planner 'minimizes the objective function (10)' (Section IV-C). The EKF covariance update (9) is independent of the measurement value when evaluated at a fixed linearization point, but the mean update and subsequent linearization depend on the sample, so the overall cost is affected. The authors provide no convergence or unbiasedness argument for using a single sample. The reader's rationale did mention this mismatch as 'the most serious technical issue,' so there is partial agreement; however, the formal weakest_assumption field points elsewhere. The proposed test—running the planner with multiple seeds and comparing true expected costs via Monte Carlo rollouts—would empirically settle whether the single-sample approximation is adequate in the tested scenario. Until this is addressed, the claim of minimizing (10) is unverified, supporting the CONDITIONAL verdict. I do not propose moving to REJECT or ACCEPT because the framework may be salvageable by either reformulating the objective or averaging over samples, as the test would clarify.","tokens_in":10439,"tokens_out":3864,"duration_ms":42858,"concrete_test":"Run the planner with multiple random seeds on the same corridor scenario, varying only the sampled observation noise in Algorithm 1. Compare the resulting plans and compute the true expected cost (10) for each plan via 1000 Monte Carlo rollouts of observation noise. If plans differ across seeds or the true expected cost exceeds the minimum by more than a small threshold, the single-sample approximation is unsound. Additionally, compare single-sample versus multi-rollout cost estimates for a fixed set of candidate actions to determine if ranking changes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the planner synthesizes a sequence of waypoint actions that minimize the objective (10), which is an expectation over future observations. However, Algorithm 1 (Section IV-D) simulates exactly one future observation at each belief update: the nominal measurement h(x, lm_i) is corrupted with a single noise draw, and the EKF update (9) is applied to that sample. No averaging over observation realizations occurs anywhere in the search. Consequently, the posterior mean (and thus any distance-to-goal term in the stage cost) depends on the particular random noise sequence, and even the covariance evolution can be affected indirectly through the linearization point at the next step. The planner therefore minimizes the cost of one randomly selected observation trajectory, not the expected cost over the observation distribution. Different random seeds can produce different plans, and the returned plan is not certified to be optimal, or even near-optimal, with respect to the stated J in (10). The paper provides no argument that a single sample is sufficient, no Monte Carlo validation, and no discussion of variance. This directly undermines the strongest claim of minimizing (10). The claim should be weakened to \"minimizes a single-observation approximation of (10)\" or the planner should average over multiple observation samples (e.g., via Monte Carlo rollouts). The reader identified this mismatch in the rationale but listed discretization as the weakest assumption; both are real, but the cost mismatch is more fundamental because it challenges the objective itself, not just the accuracy of the belief simulation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an integrated Task-Motion Planning framework for mobile-robot navigation under state uncertainty. The planner samples collision-free poses with a potential-field RRT, encodes the navigation problem in PDDL+, and uses semantic attachments inside the DiNo planner to perform EKF belief propagation and updates during a weighted A* search. The stated objective is a finite-horizon cost combining distance to the goal and the covariance trace, with an expectation over future observations. The approach is evaluated in a Gazebo corridor environment for several planner and motion discretizations, and an extended experiment adds a battery-discharge constraint. The authors claim this is the first TMP approach based on PDDL+ planning semantics.","tokens_in":10670,"tokens_out":5677,"duration_ms":65143,"significance":"If the central claims hold, the paper makes a useful integration contribution: it demonstrates that belief-space reasoning can be carried out inside a hybrid PDDL+ planner through semantic attachments, and that the resulting uncertainty estimates can guide heuristic search. The PDDL+ domain description is a reusable artifact, and the empirical study addresses a realistic navigation scenario with VAL validation and an explicit discussion of discretization effects. However, the significance is currently limited by a mismatch between the formal objective in Eq. (10) and the implemented planner, and by the single-scenario, single-sample experimental validation. The framework is plausible and worth pursuing, but the specific claim that the planner minimizes Eq. (10) is not established as written.","major_comments":[{"comment":"The objective in Eq. (10) is an expectation over future observations, but the planner replaces the observation process with a single randomized measurement: Section IV-D states that the nominal observation is corrupted with noise to obtain z_{k+1}, and Algorithm 1 applies the EKF update to that single draw. No averaging over observation realizations is described, so the generated plan minimizes the cost of one sampled observation trajectory, not the expectation in (10). This is load-bearing because the posterior mean and the linearization points for subsequent EKF steps depend on the sampled noise. The authors should either weaken the claim to a single-observation approximation of (10) and support it empirically, or average over multiple observation samples within the search.","section":"Section IV-D, Algorithm 1, Eq. (10)"},{"comment":"Even under a single-observation approximation, the paper does not specify how the full cost (10) is evaluated during search. Algorithm 1 returns only trace(Σ) in its event effects, and Section IV-B says only that cΣ is absorbed into g(·); no implementation of the distance-to-goal term cG is given, and the EKF posterior mean is not returned to the PDDL state. Since the mean affects cG and future linearizations, the implemented weighted A* search is underspecified with respect to the stated objective. The authors should provide a precise mapping from PDDL state variables and semantic-attachment effects to the g(·) and heuristic costs actually used in search.","section":"Section IV-C and IV-D, Algorithm 1"},{"comment":"The empirical evaluation does not support the quantitative claims about optimal discretization choices. Each row of Table I appears to report a single run, which is problematic because the planner uses randomized observations; a single run does not characterize the distribution of trace(Σ_g). The observation in Fig. 5c that the robot might collide with walls is based on visual inspection of a covariance curve rather than an executed collision check or path feasibility analysis. Repeating experiments over multiple random seeds and reporting summary statistics, or switching to a deterministic nominal-observation update, would be required to support statements such as 'the optimal values for Δ and dFactor are 1.00 and 2 respectively'.","section":"Section V, Table I and Fig. 5"}],"minor_comments":[{"comment":"The text says DiNo uses a modified Enforced Hill-Climbing search and then states that the search is modified to weighted A*; please clarify which search is actually used for the reported experiments.","section":"Section IV-B"},{"comment":"The word 'ratified' appears in 'the plans synthesized are ratified against the continuous model'; 'validated' or 'verified' is the intended term.","section":"Section V"},{"comment":"The phrase 'the extant of such pruning' should be 'the extent of such pruning'.","section":"Conclusion"},{"comment":"In Eq. (6), the index inside the product should be i rather than k: the product should read ∏_{i=1}^k p(x_i | x_{i-1}, u_{i-1}) p(z_i | x_i).","section":"Eq. (6)"},{"comment":"Reference [19] contains the typo 'Airtificial Intellignece'; also, Fig. 3's caption writes 'precess' where 'process' is meant.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The relationship to the authors' prior work [24] should be clarified: the paper is described as an extension, but the specific incremental contribution beyond [24] is not stated explicitly. Also, the 'first PDDL+-based TMP' claim should be checked against related semantic-attachment planners such as [1] and the navigation-focused work in [16], since the novelty claim may be easy for a reviewer to contest."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful core here is the integration pattern: route PDDL+ process/event semantics through semantic attachments so an EKF belief update runs inside a hybrid task planner, with the resulting covariance feeding the search heuristic. That is a clean way to get perception-aware navigation plans out of a planner like DiNo, and the first-PDDL+-TMP claim is credible. The EKF equations (8)-(9) are standard and correctly stated, and the potential-field RRT sampling is a sensible way to densify waypoints around landmarks. I'd read this as an incremental system contribution over the authors' own [24], not a new algorithm.\n\nThe soft spot is the objective. Equation (10) is an expectation over future observations; Algorithm 1 generates exactly one noisy measurement at each update and never averages. The planner is therefore optimizing a random observation sample path. Different random seeds can produce different plans, and there is no argument, Monte Carlo or otherwise, that the single sample is representative. This undermines the central \"minimizes J\" claim. It should either average over observation samples or be restated as an approximation. This is more fundamental than the discretization concern, though that one is real too: the paper admits a coarse Δ can skip decision points, and no adaptive step or error bound is provided.\n\nThe empirical section is too thin. One corridor scenario in Gazebo, no baselines (PETLON or a non-belief-aware planner), no error bars or repeated runs. The choice of Δ = 1.0, dFactor = 2 is made post hoc from the table, so the \"optimal\" label isn't a finding. And the \"no valid plan exists\" claim at 40% charge is really \"the planner ran out of memory after 140000 states\"; memory exhaustion is not an existence proof. The related work coverage is fine; PETLON is correctly identified as the closest comparison.\n\nWhat's here is plausible and clearly written. The belief update equations are correct; the framework is a combination of known components. If the authors fix the objective mismatch, add at least one baseline, and drop the memory-exhaustion overclaim, this could be a reasonable systems paper for an IROS/ICAPS workshop or a conference in this area. As is, I'd recommend major revision, not desk rejection: the central mechanism is worth refereeing and the flaws are fixable. Send it out, but ask for a careful response on the cost mismatch.","headline":"A plausible PDDL+-semantic-attachment integration for belief-space navigation TMP, but the planner minimizes a single noisy observation path rather than the expectation in its stated objective, and the validation is too thin to support the claims as written.","tokens_in":11282,"tokens_out":2753,"would_cite":true,"duration_ms":29568,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hybrid task-motion planner folds robot pose uncertainty into PDDL+ task planning, producing waypoint paths that minimize both travel distance and belief covariance.","keywords":["task-motion planning","belief space planning","PDDL+","semantic attachments","mobile robot navigation","partial observability","extended Kalman filter","hybrid planning"],"falsifier":"Simulate the same corridor scenario in continuous time, recording the true EKF covariance trace along the executed waypoint sequence, and compare it with the trace the planner predicted at $\\Delta = 0.5, 1, 2, 3$. If the predicted trace differs materially from the continuous-time trace, or if the planner's chosen waypoint sequence changes when $\\Delta$ is refined, the discretized belief model is not faithful.","tokens_in":10151,"feed_emoji":"🤖","tokens_out":4625,"duration_ms":42728,"temperature":0.7,"pith_summary":"The paper tries to establish that a single task-motion planner can reason about robot pose uncertainty while it plans, rather than treating task selection and motion feasibility as separate stages. It proposes an integrated framework in which a PDDL+ hybrid planner simulates motion as a continuous process and triggers belief updates through semantic attachments that call an external extended Kalman filter. If right, a robot can synthesize waypoint sequences that jointly minimize distance-to-goal and state covariance, and the same domain model works for any mobile robot navigation problem. The motivating scenario is a robot that must reach a charging station with bounded pose uncertainty and enough battery.","feed_headline":"One planner folds robot uncertainty into path choice","feed_subtitle":"Task-motion planner uses PDDL+ and Kalman-filter belief updates to find waypoint paths that reach goals with bounded pose error.","key_machinery":"The load-bearing mechanism is the PDDL+ process-event pair: a process odometry simulates the robot translation in steps of $\\delta_{\\mathrm{trans}}^k = \\Delta \\times d_{\\mathrm{factor}}$, and an event belief update triggers an external semantic attachment that computes the EKF propagated belief and, when a landmark is in range, the posterior belief, attaching the covariance trace to the event effect. These returned values guide the staged relaxed-plan-graph heuristic and the weighted A* search, so the belief estimate is not a post-hoc check but part of the search guidance. Semantic attachments, meaning external procedures attached to function symbols, are the bridge that lets a discrete task planner call numerical belief updates without leaving PDDL+ semantics.","core_discovery":"The central claim is that PDDL+ processes and events, combined with semantic attachments whose computed belief values feed the planner's heuristic, can carry belief-space reasoning inside task planning. The planner searches over a sampled set of collision-free poses, and for each candidate waypoint action it simulates the motion with a discretized odometry process; each simulation step triggers an event that calls an external library to apply EKF prediction and update equations (8)-(9), returning the trace of the covariance as an effect. The search is a weighted A* that folds the covariance trace into the cost, so the resulting plan minimizes the objective (10), a sum of distance-to-goal and uncertainty costs. The authors present this as the first TMP approach based on PDDL+ planning semantics, and they validate it in a simulated corridor where the robot must reach a charging dock.","pith_inferences":["The discretization dependence suggests a natural testable extension: an adaptive step-size controller that refines $\\Delta$ near landmarks or high-curvature belief changes, or a closed-form bound relating discretization error to covariance-trace error.","The belief-update-count plots expose a possible degeneracy: a coarse $\\Delta$ skips the very measurements that would reduce uncertainty, so the optimal-looking plan in trace terms may be invalid; a planner that reports which decision points were skipped could flag this.","The framework's separation of task search from the external EKF means the same scaffolding could attach other estimators, such as particle filters or non-Gaussian beliefs, without changing the PDDL+ model, as long as the attachment returns a scalar cost."],"forward_implications":["A robot can be given goal conditions that mix symbolic and geometric requirements, such as reaching a pose with covariance trace below a threshold, and the planner will reject action sequences that violate them.","The same PDDL+ domain description can serve any mobile robot navigation problem; only the map, landmarks, initial belief, and goal pose change.","Because belief updates happen during heuristic construction, planning time grows with the number of sampled poses and the fineness of discretization; coarser settings trade accuracy for speed.","The battery model shows task-level constraints like power limits can be folded into the same plan, so a motion-only plan can be catastrophic when it ignores them."],"supporting_citations":[{"why":"Supplies PDDL+ processes and events, the modeling constructs used to simulate continuous motion and trigger belief updates.","marker":"[7]"},{"why":"Provides the heuristic PDDL+ planner that the paper extends with semantic attachments and a weighted A* search.","marker":"[22]"},{"why":"Shows how external advisors can update indirect variables and guide heuristic construction, the basis for using attachment effects in search.","marker":"[1]"},{"why":"Defines semantic attachments for domain-independent planning, the bridge that calls the external EKF from PDDL+.","marker":"[4]"},{"why":"Is the closest navigation TMP baseline, assuming full observability and trajectory-length cost, which this paper contrasts with belief-space costs.","marker":"[16]"},{"why":"Is the prior work this paper extends, supplying the perception-aware task-motion planning idea.","marker":"[24]"},{"why":"Provides the sampling-based path generation used to build sets of collision-free poses for task planning.","marker":"[14]"},{"why":"Validates discretized plans against the continuous model, exposing the valid-but-flawed behavior for coarse discretization.","marker":"[10]"}],"fun_headline_variants":["Belief-space planning with PDDL+ and Kalman filters","Task-motion planner with uncertainty-aware pathfinding","Robot planner integrates belief updates into task planning","Hybrid planner combines belief state with task goals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Planner discretization of motion (the step size $\\Delta$ and motion factor $d_{\\mathrm{factor}}$) accurately represents the true continuous belief trajectory; the paper itself notes a coarse $\\Delta$ can skip decision points and yield valid but flawed plans.","fun_headline_variants_meta":{"raw":{"variants":["Belief-space planning with PDDL+ and Kalman filters","Task-motion planner with uncertainty-aware pathfinding","Robot planner integrates belief updates into task planning","Hybrid planner combines belief state with task goals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000169,"raw_usage":{"total_tokens":1197,"prompt_tokens":807,"completion_tokens":390,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":423,"completion_tokens_details":{"reasoning_tokens":328}},"tokens_in":423,"tokens_out":390,"duration_ms":4329,"temperature":1.0,"reasoning_tokens":328,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:50:27.954904+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the same corridor scenario in continuous time, recording the true EKF covariance trace along the executed waypoint sequence, and compare it with the trace the planner predicted at $\\Delta = 0.5, 1, 2, 3$. If the predicted trace differs materially from the continuous-time trace, or if the planner's chosen waypoint sequence changes when $\\Delta$ is refined, the discretized belief model is not faithful.","supporting_citations":[{"cited_title":"Modelling Mixed Discrete-Continuous Domains for Planning","cited_arxiv_id":null,"evidence_quote":"Supplies PDDL+ processes and events, the modeling constructs used to simulate continuous motion and trigger belief updates."},{"cited_title":"Heuristic Planning for PDDL+ Domains","cited_arxiv_id":null,"evidence_quote":"Provides the heuristic PDDL+ planner that the paper extends with semantic attachments and a weighted A* search."},{"cited_title":"Boosting Search Guidance in Problems with Semantic Attachments","cited_arxiv_id":null,"evidence_quote":"Shows how external advisors can update indirect variables and guide heuristic construction, the basis for using attachment effects in search."},{"cited_title":"Semantic Attachments for Domain-Independent Planning Systems","cited_arxiv_id":null,"evidence_quote":"Defines semantic attachments for domain-independent planning, the bridge that calls the external EKF from PDDL+."},{"cited_title":"Petlon: Planning efficiently for task-level-optimal navigation","cited_arxiv_id":null,"evidence_quote":"Is the closest navigation TMP baseline, assuming full observability and trajectory-length cost, which this paper contrasts with belief-space costs."},{"cited_title":"Towards Perception Aware Task-Motion Planning","cited_arxiv_id":null,"evidence_quote":"Is the prior work this paper extends, supplying the perception-aware task-motion planning idea."},{"cited_title":"Rrt-connect: An efficient approach to single-query path planning","cited_arxiv_id":null,"evidence_quote":"Provides the sampling-based path generation used to build sets of collision-free poses for task planning."},{"cited_title":"V AL: Automatic plan validation, continuous effects and mixed initiative planning using PDDL","cited_arxiv_id":null,"evidence_quote":"Validates discretized plans against the continuous model, exposing the valid-but-flawed behavior for coarse discretization."}],"review_version":1}