{"id":"fa92df9e-81d6-425b-86bf-d7fee038ec20","arxiv_id":"2411.12837","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An anticipatory task planner with a GNN over 3D scene graphs reduces long-horizon task costs by 5 to 42 percent in simulated home and restaurant settings.","lead":"This paper describes a robot planner that looks ahead to future tasks instead of only optimizing the current one, using a learned graph-based model to estimate later costs. In simulated homes and restaurants, the authors report 5.38 to 42.5 percent lower total task-sequence costs, with the largest gains when the robot can prepare the environment in advance.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Estimator accuracy and train/eval separation are unverified: the paper trains and evaluates on the same task distribution P(τ), reports no held-out map split, no error bars, and no oracle baseline.","rationale":"The paper's central claim is an empirical one, and I read it in that spirit: if the estimator is accurate and the evaluation is properly held out, the reported cost reductions are meaningful. I do not find a logical inconsistency in Eq. (1) or the sampling procedure. The most defensible concern is that the evidence for the estimator's accuracy is incomplete. Section IV-C explicitly trains on tasks sampled from P(τ), and Section V evaluates on tasks drawn from the same distribution; the conclusion acknowledges this as a limitation for unseen tasks, but the paper does not state whether the evaluation maps are held out from training. Because preparation directly optimizes under P(τ), a memorizing estimator could inflate the 40.6% and 42.5% preparation numbers. In addition, Table I gives only point estimates; the 5.38% home gain is small and no uncertainty is reported. These are not accusations of misconduct; they are missing supports that a reproducibility check can settle. I therefore retain the conditional verdict.","tokens_in":10220,"tokens_out":7707,"duration_ms":87782,"concrete_test":"Obtain the raw per-sequence costs and the map IDs for training and evaluation. Recompute Table I after excluding any evaluation home or restaurant whose layout was used to generate training data, enforcing a strict held-out split, and report paired 95% bootstrap confidence intervals and a paired significance test for each row. If the 5.38% home gain or the 40.6% preparation gain loses significance or vanishes under this split, the central claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim rests on the learned expected-cost estimator being accurate on the evaluation distribution. Two benchmark-validity gaps make this insecure. First, Section IV-C generates training labels from the same task distribution P(τ) that Section V uses for evaluation, and the paper never states whether the 500 evaluation PROCTHOR homes and the restaurant layouts are disjoint from the maps used to generate training data. If they overlap, the reported gains, especially the 40.6% preparation gain, which directly optimizes under P(τ), could reflect layout-specific memorization rather than a generalizable estimator. Second, Table I reports only point estimates over 500 sequences with no error bars, confidence intervals, or paired significance tests, so the 5.38% home improvement is not yet distinguishable from run-to-run noise. The limitation stated in the conclusion, that the robot may encounter unseen tasks, is real but applies to deployment; the benchmark-internal holdout and statistical questions must be settled before the headline reductions are treated as established. An oracle baseline or an estimator-error metric is also missing, so the specific contribution of the learned estimator versus the sampling procedure is not isolated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a learning-augmented, model-based anticipatory task planning framework for large-scale domestic environments. It learns a GNN-based estimator of expected future cost from a 3D scene graph representation and combines it with focused sampling of augmented tasks for anticipatory planning and simulated annealing for preparation. Experiments in PROCTHOR/ALFRED homes and a custom restaurant domain report average cost reductions of 5.38% and 31.5% over myopic planning, and 40.6% and 42.5% when preparation is allowed. The central claim is that the learned estimator plus focused plan sampling yields cheaper long-horizon task execution than planning only for the immediate task.","tokens_in":10455,"tokens_out":5198,"duration_ms":50101,"significance":"If established, the result would be a useful step toward long-horizon task planning in persistent environments, where myopic planners are known to suffer from side effects. The evaluation has real strengths: it uses a standard PDDL planner (FastDownward), spans 500 PROCTHOR homes and 500 restaurant layouts with 10,000 and 20,000 task executions, and includes a preparation mode that goes beyond anticipatory replanning. The paper also builds on a previously published formulation rather than inventing a new objective. However, the statistical and generalization evidence for the headline numbers is incomplete, and the learned estimator's accuracy is not directly measured; these gaps are load-bearing for the paper's central claim.","major_comments":[{"comment":"The reported reductions are point estimates over 500 sequences per environment, with no standard deviations, confidence intervals, or paired significance tests. The 5.38% improvement in PROCTHOR is small relative to the variability one would expect across 500 randomly generated homes, so it is not yet distinguishable from run-to-run noise. The authors should report the per-sequence pairwise difference between methods and a paired test (e.g., Wilcoxon signed-rank) across the 500 sequences, for both environments.","section":"Section V, Table I"},{"comment":"The paper never states whether the evaluation maps are disjoint from the maps used to generate training data. Both training and evaluation tasks are drawn from the same distribution P(tau), and Section V-B says 'The tasks for data generation and evaluation are taken from this task distribution,' which leaves open the possibility that the same homes or restaurant layouts appear in both sets. If there is overlap, the reported gains, especially the 40.6% preparation gain that optimizes under P(tau), could reflect layout-specific memorization rather than generalization. An explicit map-level train/eval split with results on held-out layouts is needed to support the claim that the estimator transfers.","section":"Sections IV-C, V-B, V-C"},{"comment":"No estimator-accuracy metric is reported anywhere. The APCost estimator is the learned component that drives both anticipatory planning and preparation, but the paper does not give its MAE, R^2, or correlation against ground-truth expected cost on a validation set. Without such a metric, the reader cannot tell whether the learned model is actually accurate, or whether the downstream planning improvements are caused by the estimator at all.","section":"Sections V-A, V-B, V-C"},{"comment":"The contribution of the learned estimator is not isolated from the sampling procedure. The anticipatory planning pipeline consists of focused sampling of augmented tasks plus the APCost estimator, and the preparation pipeline uses simulated annealing with the same estimator. An ablation with an oracle estimator (or a sampling-only baseline) is needed to establish that the learned estimator, rather than the broader search over plans, is responsible for the reported cost reductions.","section":"Sections IV-A, V"},{"comment":"The conclusion explicitly acknowledges that the robot may encounter unseen tasks outside P(tau), which is a real deployment limitation. However, the benchmark-internal question is more basic: even within P(tau), the paper does not demonstrate that the estimator generalizes to held-out maps. That internal generalization issue should be settled before the headline reductions are treated as established; the unseen-task limitation alone is not a substitute for a train/eval split in the experiments.","section":"Conclusion"}],"minor_comments":[{"comment":"The spacing in 'PROCTHOR' is inconsistent; the text uses 'P ROCTHOR' in many places (e.g., Fig. 1, Section V-B) and 'PROCTHOR' elsewhere.","section":"Throughout"},{"comment":"The task name 'ClearConatiners' appears to be a typo for 'ClearContainers.'","section":"Section V-C"},{"comment":"The term 'APC OST ESTIMATOR' should read 'APCost Estimator' for readability and consistency with the text.","section":"Section IV-A, Algorithm 1"},{"comment":"The phrase 'employing myopic planning insignificantly reduces the average cost per task over time' is ambiguous; please clarify whether the small decrease is statistically insignificant and what 'insignificantly' means here.","section":"Section V-B"},{"comment":"The restaurant environment is of the authors' design and is not a public benchmark; this limits comparability with prior work unless the environment and task distributions are released.","section":"Section V-C"},{"comment":"Equation (1) minimizes the immediate cost plus the expected cost of a single next task, while the evaluation measures cost over 20-40-task sequences; a short discussion of why one-step anticipation is sufficient for long-horizon improvement would help.","section":"Section III-B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript would benefit from a clearer statement of its incremental contribution over the authors' prior ICRA 2023 paper [10], since the formulation and Eq. (1)-(2) are carried over directly. The absence of any baseline beyond myopic planning, and the lack of estimator-error and statistical evidence, makes the quantitative claims hard to verify. I would be more comfortable recommending acceptance if the authors supply held-out-map results, paired significance tests, and an oracle/sampling ablation; these are within the scope of the current experimental setup."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a reasonable step from the authors' earlier blockworld anticipatory planning work to large-scale ProcTHOR and restaurant domains. The new pieces are the GNN over a 3D scene graph to estimate expected future cost and a focused sampling scheme that augments the current task with extra predicates in a bounded region. That combination is sensible, and the evaluation is large by the standards of this subfield: 500 sequences in each domain, 10k and 20k task executions. The restaurant numbers look genuinely strong—31.5% cost reduction without preparation, 42.5% with preparation. The home improvement of 5.38% is plausible but too small to take seriously without error bars.\n\nThe soft spots are mostly about evidence quality. There are no confidence intervals or paired significance tests anywhere, so the home gain is within the range of run-to-run noise. More importantly, the training and evaluation tasks come from the same distribution P(tau), and the paper never says whether the 500 evaluation homes and restaurant layouts are disjoint from the ones used to generate training data. If they overlap, the gains—especially the 40.6% preparation gain, which optimizes directly under P(tau)—could partly reflect memorization rather than generalization. The authors do state the distribution assumption up front and acknowledge in the conclusion that real robots may see unseen tasks. That is honest, but it doesn't settle the benchmark-internal holdout question. I also would have liked an oracle baseline or at least an estimator-error metric (e.g., MAE on held-out states) to tell how much of the gain comes from the learned estimator versus the sampling procedure. The restaurant environment is custom, so it isn't a public benchmark, though the ProcTHOR portion does use a standard base.\n\nThere's also the self-reference point: the formulation follows the authors' own prior paper [10]. That is not a problem by itself, and they cite it properly. The new contribution here is genuinely the scaling, not the formulation.\n\nBottom line: the method is coherent, the direction is credible, and the restaurant results would be a real contribution if the statistics hold. This paper deserves a proper peer review, but I'd push for revision—error bars, a train/eval map split, and estimator accuracy numbers—before accepting the headline percentages. If I worked in task planning, I'd cite this as a proof-of-concept for learning-based anticipatory cost at scale, with a note that the quantitative claims are not yet established. Bring it to reading group if you want to discuss how to evaluate learned cost models in simulated homes.","headline":"A credible scaling of anticipatory planning to large simulated homes and restaurants, but the headline gains rest on point estimates and same-distribution train/test, so treat them as promising rather than proven.","tokens_in":10974,"tokens_out":2549,"would_cite":true,"duration_ms":25438,"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":"The paper claims that a graph-neural-network anticipatory planner reduces long-horizon task-sequence costs by 5.38% in homes and 31.5% in restaurants, and by 40.6% and 42.5% respectively when the robot can prepare the environment in…","keywords":["anticipatory planning","long-horizon task planning","graph neural networks","3D scene graphs","service robots","PDDL task planning","environment preparation","task sequence cost"],"falsifier":"Take the trained estimator and planner and evaluate on task sequences drawn from a distribution that intentionally excludes all training tasks but stays within the same environment layouts—for example, novel object categories or new cleaning and serving tasks. If the anticipatory planner's average sequence cost is not below the myopic baseline, or if preparation no longer reduces cost, the generalization claim behind the reported numbers is falsified. A more gradual version is to measure whether the cost advantage decays monotonically as the test distribution is perturbed away from the training distribution.","tokens_in":10044,"feed_emoji":"🤖","tokens_out":8580,"duration_ms":74589,"temperature":0.7,"pith_summary":"The paper sets out to show that a service robot working through a long sequence of household or restaurant tasks can plan better by anticipating the expected cost of future tasks, not just the immediate task at hand. It claims that myopic planning—choosing the cheapest way to finish today's task—can make later tasks more expensive, while slightly costlier choices now can lower total cost across the sequence. To make this anticipatory planning tractable in large, realistic environments, the paper combines a graph neural network trained on a compact scene-graph-like representation to estimate a state's expected future cost with a focused sampling procedure that generates candidate plans. The reported result is that this approach reduces average task-sequence cost by 5.38% in homes and 31.5% in restaurant settings, and by 40.6% and 42.5% respectively when the robot is allowed to prepare the environment in advance.","feed_headline":"Robot planner anticipates future tasks, cutting costs by up to 42.5%","feed_subtitle":"A GNN that estimates future task costs lets service robots choose actions that make the next jobs cheaper.","key_machinery":"The mechanism is a learned expected-cost estimator paired with a focused plan sampler. The estimator is a graph neural network whose input is the environment encoded as a graph: nodes are semantic entities (rooms, containers, objects) with features for type, location, and state, and edges encode spatial and logical relations; a sequence of graph-convolution layers with pooling maps each graph to $V_{A.P.}(s)$, the expected cost of completing a follow-up task from that state. On the planning side, candidate plans are generated by solving 'augmented tasks'—the current task extended with additional predicates on entities within a bounded region of the robot's path—using a classical symbolic planner. The planner then keeps the plan whose terminal state has the lowest sum of immediate cost plus estimated future cost. Preparation uses the same estimator inside a simulated-annealing search over environment states.","core_discovery":"The paper's central claim is that anticipatory planning, which minimizes the sum of immediate task cost and the expected cost of the next task drawn from a known distribution, can be scaled from small block-world domains to large home-like environments and that doing so materially reduces long-horizon costs. The concrete instantiation is a graph neural network over a scene-graph-like representation that predicts the anticipatory planning cost $V_{A.P.}(s)$ of any state, integrated with a planner that samples augmented tasks—the current task plus extra predicates involving nearby entities—and selects the plan whose terminal state minimizes $V_{s'_g}(s_0) + V_{A.P.}(s'_g)$. The same estimator powers a preparation phase that searches for a low-expected-cost state before any task is assigned, using simulated annealing. In randomized large-scale home and restaurant environments, the anticipatory planner beats the myopic baseline on average cost per task, and preparation yields the largest gains.","pith_inferences":["If the estimator generalizes, the same learned expected-cost function could serve as a heuristic for other long-horizon problems, such as multi-robot coordination or persistent human-robot collaboration, where today's actions create side effects for later activities.","A testable extension would train the estimator on task distributions with different long-tail structures and measure how gracefully preparation gains degrade; the paper's known-distribution assumption makes this the main open risk.","The home gain of 5.38% likely understates the approach's potential because the home tasks used are essentially pick-and-place; re-running the same framework with cleaning, filling, or washing tasks in homes, as was done in the restaurant domain, would probably widen the gap.","The paper's preparation result implies a practical scheduling rule: when the robot has free time, it should spend it rearranging the environment toward lower expected cost rather than idling, provided the task distribution is stable."],"forward_implications":["A service robot can reduce total task-sequence cost using only the task distribution, not knowing exactly which task comes next, by accepting slightly higher immediate cost for a state that makes likely follow-up tasks cheaper.","Idle time before a task is assigned is valuable: guided by the learned cost estimate, even a myopic planner becomes substantially cheaper when the environment is prepared in advance.","Anticipatory planning remains tractable as environments grow because the focused sampling procedure only needs to explore augmented tasks involving entities near the robot's path, not all possible completions.","The larger gains in restaurant settings than in homes suggest that the approach's benefit grows with task interdependency, since placement-only tasks leave fewer side effects to exploit."],"supporting_citations":[{"why":"formalizes the anticipatory planning objective and the preparation formulation that this paper scales up","marker":"[10]"},{"why":"supplies the procedurally generated home environments used for training and evaluation","marker":"[13]"},{"why":"provides the pick-and-place task distribution for the home experiments","marker":"[14]"},{"why":"is the classical planning system used to compute immediate plan costs and solve augmented tasks","marker":"[19]"},{"why":"introduces the 3D scene graph representation that inspires the graph encoding of the environment","marker":"[11]"},{"why":"defines the graph network framework underlying the expected-cost estimator","marker":"[12]"},{"why":"provides the transformer-based graph convolution layer used in the estimator","marker":"[38]"},{"why":"supplies the simulated annealing procedure used for environment preparation","marker":"[34]"}],"fun_headline_variants":["GNN anticipates future tasks, cutting robot costs by up to 42.5%","Anticipatory planner reduces robot costs 5-42% across homes and restaurants","Scene-graph GNN enables anticipatory planning for long-lived robots","Robot prep: GNN estimates future costs, cutting task sequences by 42.5%","Anticipatory planning with GNN: up to 42.5% lower cost for robots"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that the robot knows the task distribution $P(\\tau)$ and that both training and evaluation tasks are drawn from it; if deployed tasks fall outside that distribution, the learned cost estimates become unreliable and the reported gains, especially the preparation gains, may fail to appear. The paper itself acknowledges this assumption in its conclusion.","fun_headline_variants_meta":{"raw":{"variants":["GNN anticipates future tasks, cutting robot costs by up to 42.5%","Anticipatory planner reduces robot costs 5-42% across homes and restaurants","Scene-graph GNN enables anticipatory planning for long-lived robots","Robot prep: GNN estimates future costs, cutting task sequences by 42.5%","Anticipatory planning with GNN: up to 42.5% lower cost for robots"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001128,"raw_usage":{"total_tokens":4699,"prompt_tokens":968,"completion_tokens":3731,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":3622}},"tokens_in":584,"tokens_out":3731,"duration_ms":24134,"temperature":1.0,"reasoning_tokens":3622,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:09:09.938050+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained estimator and planner and evaluate on task sequences drawn from a distribution that intentionally excludes all training tasks but stays within the same environment layouts—for example, novel object categories or new cleaning and serving tasks. If the anticipatory planner's average sequence cost is not below the myopic baseline, or if preparation no longer reduces cost, the generalization claim behind the reported numbers is falsified. A more gradual version is to measure whether the cost advantage decays monotonically as the test distribution is perturbed away from the training distribution.","supporting_citations":[{"cited_title":"Anticipatory planning: Improving long-lived planning by estimating expected cost of future tasks,","cited_arxiv_id":null,"evidence_quote":"formalizes the anticipatory planning objective and the preparation formulation that this paper scales up"},{"cited_title":"The Fast Downward planning system,","cited_arxiv_id":null,"evidence_quote":"is the classical planning system used to compute immediate plan costs and solve augmented tasks"},{"cited_title":"3D Scene Graph: A structure for unified semantics, 3D space, and camera,","cited_arxiv_id":null,"evidence_quote":"introduces the 3D scene graph representation that inspires the graph encoding of the environment"}],"review_version":1}