{"id":"b3e48873-a56a-4e0b-ae1d-c51fe4ac41b9","arxiv_id":"2411.17326","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A flat POMDP model with a known stochastic worker policy lets an online MCTS planner with goal-based reward shaping outperform plain POMCP on two simulated active goal recognition tasks.","lead":"This paper models a robot assistant that anticipates a human worker's next actions and chooses helpful actions such as perceiving, inspecting, and restocking under noisy observations. It compares two online POMDP planners on two simulated factory tasks and reports that its relevance-based planner achieves higher returns than a standard baseline.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Formal model in §4 defines target transitions as π_T(s_T) only, so observer actions cannot affect target state; but the experiments require the opposite, leaving the model as written inconsistent with the reported results.","rationale":"The reader's weakest assumption (known π_T and G_T) is a real limitation, but the paper explicitly frames the problem that way in Definition 1, so it is a boundary of the proposal rather than an internal contradiction. The more serious issue is that the model as formally specified is inconsistent with the experiments. If target transitions depend only on s_T, then no observer action can change the target's ability to reach G_T; yet all reward structure in §5 is driven by the robot providing tools or parts (e.g., +5 if a tool is brought and needed, and the worker cannot succeed without the robot). The intended model must let π_T depend on observer-controlled variables, but the paper never introduces this dependence. This makes the central claim that active intention recognition tasks can be modeled by the tuple <Σ, s0, π_T, G_T> unsupported by the written formalism. The concrete check would either confirm the model is degenerate or force a revision of the transition definition. Because the fix is straightforward and the experimental idea may still be valid, the reader's CONDITIONAL verdict stands, but for a different reason. I therefore agree partially with the reader: same verdict, different load-bearing issue.","tokens_in":29,"tokens_out":8152,"duration_ms":137647,"concrete_test":"Re-implement the maintenance scenario transition function literally from the §4 definition, with s'_T sampled from π_T(s_T) and observer actions affecting only S_O. Check whether the robot's \"bring tool\" action changes the worker's probability of successfully inspecting or replacing components. If it does not, the model cannot produce the reported reward structure; if it does, the formal definition must be revised to make π_T (or the target transition) a function of the full state s = (s_O, s_T). In the absence of released code, this re-derivation from the domain description in §5.1 is the decisive check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 4, the paper defines the POMDP with state space S = S_O × S_T and states that \"the transition model T = T(s, a, s')π_T(s_T), such that Target variables s'_T = s' ∩ S_T are determined by π_T(s_T)\". Under this definition, the target's next state depends only on the current target state, so observer actions a cannot influence any target variable. Yet the experiments are explicitly designed around the assistant changing target-relevant state: in the maintenance scenario \"the worker depends on the robot's intervention to provide tools and a spare relay\", and in assembly the robot's restock actions alter container contents that determine whether the worker's assemble action succeeds. Therefore, either the implementation contradicts the formal model by letting π_T (or the overall transition) depend on observer-controlled state, or the model is degenerate and the reported reward structure is impossible. The paper never introduces the necessary coupling, and no code is released to resolve the ambiguity. This is more fundamental than the known-goals assumption: even with a perfectly known π_T and G_T, the model as written cannot represent the assistance loop that the experiments claim to evaluate.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a partially observable Markov decision process (POMDP) formulation for active intention recognition, in which a target agent (a human worker) is simulated by a stochastic model π_T and an observer (a robot) plans online with Monte-Carlo tree search. The authors introduce RAGE, a relevance-based planner that uses Partial Goal Satisfaction (PGS) reward shaping and a goal-driven rollout policy, and compare it with standard POMCP on two simulated domains: a maintenance task and a toy-truck assembly task. Across 100 episodes, RAGE reports higher average discounted returns and lower standard errors than POMCP in both domains. The authors conclude that active intention recognition tasks can be modeled and solved online with their proposal and that relevance-based planning can improve performance without highly specialized planners or extensive domain knowledge.","tokens_in":11839,"tokens_out":9109,"duration_ms":81749,"significance":"If the results hold, the paper contributes a practical, flat POMDP formulation for active goal recognition that avoids factored representations and plan libraries, and it provides an initial demonstration that the RAGE planner can outperform POMCP on two simulated assistant tasks. The authors are appropriately cautious in calling the results preliminary, and they explicitly state the limitation that the IRE component was not thoroughly tested. However, the significance is substantially weakened by (i) an inconsistency between the formal transition model and the experimental setup, (ii) the unablated reward shaping in the RAGE comparison, and (iii) missing reproducibility details. The conceptual direction — using a generative target model and online POMDP planning for active intention recognition — is timely and worth further development.","major_comments":[{"comment":"The formal transition model states that target variables s'_T = s' ∩ S_T are determined by π_T(s_T), making the target's evolution independent of the observer's actions and observer-controlled state. Yet the experiments in Sections 5.1 and 5.2 rely on the opposite: the worker 'depends on the robot's intervention to provide tools and a spare relay' (5.1) and 'the robot must continually monitor both the environment and the worker activities' because the worker 'may be unable to complete their task without the robot's help' (5.2). The paper never specifies a coupling mechanism by which observer actions affect s_T or the target's ability to achieve G_T. This is a load-bearing inconsistency: either the implementation uses a different transition model than the one formally defined, or the proposed model cannot represent the assistance loop that the experiments claim to evaluate. Please correct the transition model to include the necessary dependence of target transitions on observer state and actions, and provide code or a precise implementation specification to confirm that the model and experiments match.","section":"Section 4, transition model"},{"comment":"The reported advantage of RAGE over POMCP is confounded by the PGS reward-shaping term F = γφ(h_{t+1}) − φ(h_t) with φ(h) = α p(h). As the paper states in 5.3, this shaping is applied only during planning and simulation, not during execution, while POMCP is run with the original reward signal. The improvement attributed to relevance-based planning may instead be caused by this internal, non-executed reward bonus. The paper should include an ablation in which POMCP is given the same shaping during planning, and should report the values of α and the PGS point assignments. Additionally, the text in 4.1 says the IRE component was 'incorporated but did not thoroughly test'; please state explicitly whether IRE was active in the RAGE runs that produced Table 1, since an untested component could influence the reported results.","section":"Section 4.1 and Section 5.3"},{"comment":"The conclusion that the approach avoids 'large amounts of domain knowledge' is undercut by the design of p(h), which assigns hand-chosen points to the goal conditions G_T given in advance, and by the scaling factor α that must be tuned per problem. The paper does not describe how the point values and α were selected, nor does it report sensitivity to these choices. This is not a logical circularity in the sense of using the hidden goal to build the model, because G_T is part of the problem specification, but the claim about minimal domain knowledge is unsupported without evidence that performance is robust to reasonable variations in these hand-tuned parameters. Please either temper the claim or provide a sensitivity analysis.","section":"Section 4.1 and Section 6"},{"comment":"The experimental section omits numerous parameters needed to reproduce the numerical results: exact transition probabilities for the target models, sensor accuracies, the values of expertise levels in the reward distributions, the discount factor γ, MCTS exploration constants, particle counts, and the precise '3 types of random variations' used to prevent particle deprivation in 5.2. No code is released. These details are essential for verifying the 100-episode averages and standard errors in Table 1 and Figures 1–2. Please provide a complete experimental specification or release the simulator and planner code.","section":"Section 5"}],"minor_comments":[{"comment":"The word 'Through' is misspelled as 'Throu gh' in the title on the first page.","section":"Page 1 (title)"},{"comment":"The transition rule 'T = T(s,a,s')π_T(s_T)' is notationally ambiguous; please define whether this denotes a product of probabilities and specify how the observer and target transitions are combined.","section":"Section 4"},{"comment":"π_T is called a 'policy' but is defined as a stochastic process with a transition function T(s_T, ω_T, s'_T); please clarify whether π_T outputs actions or next states.","section":"Section 4"},{"comment":"The sentence 'Visual compartment inspections succeed with probability, as do using the screwdriver but the multimeter always returns the correct relay status' is incomplete; a concrete probability or formula is needed.","section":"Section 5.1"},{"comment":"The worker simulation is described as a deterministic loop over 'assemble' actions; please state where stochasticity enters the target model besides part availability.","section":"Section 5.2"},{"comment":"Starting the performance plots at 32 simulations for maintenance and 256 for assembly can obscure failures at lower planning budgets; please report the full curves or justify the truncation.","section":"Section 5.3"},{"comment":"The paper reports standard errors but no hypothesis tests; a paired or two-sample test across the 100 episodes would substantiate the wording 'significant improvements.'","section":"Table 1"},{"comment":"The statement that 'recent papers' recommend online POMDP planning for AGR is supported only by references [11] and [1]; please expand or rephrase to make the attribution precise.","section":"Section 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads like an extended abstract or workshop paper, and the authors are honest about the preliminary nature of the results, which is to their credit. However, the formal transition-model inconsistency and the unablated comparison must be resolved before publication in a journal. The paper's contribution is potentially useful but currently requires substantial technical revision and a more complete experimental report."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. This is a clearly written preliminary report, but the formal model and the experiments don't line up, and that's the first thing to fix.\n\nWhat's new: they cast active goal recognition as a flat POMDP with a generative target simulator, and they test RAGE, their relevance-based planner, against vanilla POMCP on two small domains. The results show RAGE getting better discounted return and lower variance, with honest reporting of standard errors and limitations. They don't oversell; the paper explicitly calls itself preliminary.\n\nThe main problem is in Section 4. The transition model defines target variables s'_T as determined only by π_T(s_T), meaning the target evolves independently of the observer's state and actions. But the experiments require the opposite: the worker's success depends on tools the robot brings, and restocking changes container contents that determine whether assembly succeeds. The paper's own prose says 'the worker depends on the robot's intervention.' So the math and the experiments are inconsistent. Without released code, a reader can't see how they resolved it. This is more serious than the known-goals assumption, because even with a perfect π_T, the model as written can't represent the assistance loop.\n\nThe known-goals assumption is a limitation they acknowledge. The bigger confound is that RAGE's advantage comes from PGS reward shaping defined directly on the given G_T, so it has domain-specific information that POMCP doesn't get. The performance gap is partly built into the comparison. A second baseline (e.g., POMCP with the same shaping) would clarify that.\n\nI'd send this to peer review because the direction is interesting and the issues are fixable, but the authors need to reconcile the formal model with their experiments and strengthen the evaluation. For my own work, I wouldn't cite it as a result yet, but I'd note it as related work.","headline":"Promising preliminary framework for online active goal recognition, but the formal POMDP model as written contradicts the experiments it reports.","tokens_in":12332,"tokens_out":7689,"would_cite":false,"duration_ms":71818,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T40","68T37"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes a flat POMDP model for online active intention recognition and shows that relevance-based planning improves on standard MCTS in two simulated robotic-assistant domains.","keywords":["partially observable Markov decision process","active goal recognition","intention recognition","online planning","Monte Carlo tree search","relevance-based planning","robotic assistant"],"falsifier":"Run simulated episodes in which the target deviates from $\\pi_T$, for example by switching goals mid-episode, choosing a different plan, or changing its action probabilities over time, and check whether the observer's policy still restocks the correct parts and reaches positive terminal returns. If RAGE's return drops to POMCP levels or the terminal state becomes unreachable in these runs, the central claim fails. A second direct test is to vary sensor noise and worker expertise systematically; if the RAGE advantage reverses or disappears under moderate parameter changes, the claimed improvement is not robust.","tokens_in":11345,"feed_emoji":"🤖","tokens_out":5308,"duration_ms":48221,"temperature":0.7,"pith_summary":"This paper tries to establish that active intention recognition, where a robotic assistant must infer and support a human worker's goals from noisy observations while doing its own tasks, can be modeled as a single flat POMDP and solved online with Monte Carlo tree search. The target worker is replaced by a small stochastic simulator $\\pi_T$ that generates actions and reward signals, avoiding the need for a hand-built plan library or a factored state representation. The paper further claims that adding relevance-based planning, specifically the goal-proximity scoring component of the RAGE planner, improves online performance over standard POMCP in two simulated industrial scenarios, yielding higher average discounted return, lower variance, and fewer simulations needed to reach the goal. If these results hold, robotic assistants could be built with less domain-specific engineering and still anticipate a human's needs in real time.","feed_headline":"Goal-proximity scoring helps robots anticipate a worker's needs","feed_subtitle":"A flat POMDP with a stochastic target model lets an assistant infer goals online, with less variance and fewer simulations.","key_machinery":"The central object is the online active intention recognition tuple $\\langle \\Sigma, s_0, \\pi_T, G_T \\rangle$, in which $\\pi_T$ is a stochastic target simulator that acts as a generative replacement for a plan library. The POMDP is kept flat: states are the product of observer and target variables, observations combine target activities and ordinary sensor readings, and the transition model is $T(s,a,s')\\pi_T(s_T)$ with reward $R = R(s,a,s') + R(s_T,a,s'_T)$. The performance improvement comes from the PGS reward-shaping function $F(h_t,h_{t+1}) = \\gamma\\varphi(h_{t+1}) - \\varphi(h_t)$ with $\\varphi(h) = \\alpha p(h)$, plus the rollout policy $A(h) = \\arg\\max_a p(h \\cup \\{a, \\omega\\})$, which directs simulation toward histories that meet goal conditions.","core_discovery":"On its own terms, the paper claims that active goal and intention recognition problems can be represented as a flat POMDP whose state includes both observer and target variables, where the target follows a known stochastic policy $\\pi_T$ and the observer maximizes discounted return, and that solving this POMDP online with MCTS-style planners works. The additional finding is that augmenting planning with partial goal satisfaction (PGS) reward shaping, which assigns points to transitions that satisfy target goal conditions $G_T$, yields significantly better and more stable returns than uniformly random MCTS in the maintenance and assembly domains, including an assembly state space larger than $10^{11}$ states. The paper is careful to call these results preliminary and notes that the incremental refinement (IRE) component of the relevance-based approach was incorporated but not thoroughly tested.","pith_inferences":["The paper assumes $\\pi_T$ is fixed and correct; a natural extension is an online belief over possible target policies, allowing the assistant to detect when the worker's behavior deviates from the modeled plan and revise its assistance accordingly.","PGS's shaping bonus may transfer to other delayed-reward human-robot collaboration tasks, such as eldercare or logistics, where the reward signal arrives only after the human completes a dependent step.","Since IRE was not thoroughly tested, a direct ablation experiment separating PGS and IRE could identify which mechanism drives RAGE's advantage and whether it persists under sensor-noise and worker-expertise variations.","The authors' future-work idea of combining POMDP goal reasoning with hierarchical task-network planning suggests a concrete test: measure whether decomposing \"bring tool\" into platform-specific actions preserves the planning advantage seen in simulation."],"forward_implications":["Active intention recognition can be tackled as ordinary online POMDP planning without factored state representations or hand-coded plan libraries.","A stochastic generative target model is sufficient to encode the human worker's behavior in these domains, with goal conditions supplied in advance.","Relevance-based planning, specifically PGS reward shaping, improves both average discounted return and variance relative to POMCP in both tested domains.","With a sufficient simulation budget, both planners reach the terminal state in the larger assembly domain, indicating scalability to state spaces over $10^{11}$.","The observer's action preferences can be generated online from goal-proximity scoring rather than hand-made rules."],"supporting_citations":[{"why":"Establishes the factored active goal recognition POMDP representation that this paper simplifies into a flat model and extends with online planning.","marker":"[1]"},{"why":"Introduces the relevance-based planning approach with PGS and IRE that RAGE implements and that produces the reported performance gains.","marker":"[20]"},{"why":"Presents incremental refinement and efficiency results for relevance-based planning, forming the basis of the RAGE planner's online action selection.","marker":"[19]"},{"why":"Defines POMCP, the Monte Carlo tree search baseline against which RAGE's performance is compared in the experiments.","marker":"[22]"},{"why":"Recommends an online planning approach for active goal recognition to avoid combinatorial explosion, motivating the paper's choice of online POMDP planning.","marker":"[11]"}],"fun_headline_variants":["Reward shaping boosts online POMDP goal prediction","Robots anticipate human actions using POMDP planning","Active goal recognition with online POMDP solvers","Partial goal rewards stabilize robot intention inference","Online POMDP planning for robotic intention recognition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The observer must know the worker's stochastic policy $\\pi_T$ and goal conditions $G_T$ in advance, and the worker must actually follow them. If the worker deviates, changes goals, or learns over time, the model has no mechanism to detect or correct the mismatch, and the claimed recognition capability collapses.","fun_headline_variants_meta":{"raw":{"variants":["Reward shaping boosts online POMDP goal prediction","Robots anticipate human actions using POMDP planning","Active goal recognition with online POMDP solvers","Partial goal rewards stabilize robot intention inference","Online POMDP planning for robotic intention recognition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000346,"raw_usage":{"total_tokens":1829,"prompt_tokens":812,"completion_tokens":1017,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":428,"completion_tokens_details":{"reasoning_tokens":945}},"tokens_in":428,"tokens_out":1017,"duration_ms":7528,"temperature":1.0,"reasoning_tokens":945,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:13:23.556876+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run simulated episodes in which the target deviates from $\\pi_T$, for example by switching goals mid-episode, choosing a different plan, or changing its action probabilities over time, and check whether the observer's policy still restocks the correct parts and reaches positive terminal returns. If RAGE's return drops to POMCP levels or the terminal state becomes unreachable in these runs, the central claim fails. A second direct test is to vary sensor noise and worker expertise systematically; if the RAGE advantage reverses or disappears under moderate parameter changes, the claimed improvement is not robust.","supporting_citations":[{"cited_title":"Planning Under Uncertainty Through Goal- Driven Action Selection","cited_arxiv_id":null,"evidence_quote":"Introduces the relevance-based planning approach with PGS and IRE that RAGE implements and that produces the reported performance gains."},{"cited_title":"Efficient planning under uncertainty with in- cremental refinement","cited_arxiv_id":null,"evidence_quote":"Presents incremental refinement and efficiency results for relevance-based planning, forming the basis of the RAGE planner's online action selection."},{"cited_title":"Monte-Carlo Planning in Large POMDPs","cited_arxiv_id":null,"evidence_quote":"Defines POMCP, the Monte Carlo tree search baseline against which RAGE's performance is compared in the experiments."},{"cited_title":"Active goal recognition using intention aware motion plan- ning","cited_arxiv_id":null,"evidence_quote":"Recommends an online planning approach for active goal recognition to avoid combinatorial explosion, motivating the paper's choice of online POMDP planning."}],"review_version":1}