{"id":"588880a7-f414-4ad2-9cf6-64ddc79e3ad9","arxiv_id":"2505.04565","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"BNG-IRL segmentation and GMR-based anomaly detection let robots learn multi-step contact tasks from a few demonstrations, detect deviations, and incrementally add recovery behaviors.","lead":"This paper presents a robot learning framework that splits a demonstration into skills using inverse reinforcement learning and clustering, then uses the learned skill constraints to detect execution errors and learn recovery behaviors. It aims to let robots learn contact-rich manipulation tasks with only a few human demonstrations on the real system.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BNG-IRL's Q-value computation is never specified: no transition model or action space is defined, so the optimality score in Eq. (4) is not computable as stated.","rationale":"The central claim rests on the segmentation step. Without a specified way to compute Q_Td and V*, the claimed unsupervised segmentation and the subsequent anomaly detection that uses the inferred skill boundaries cannot be verified. The reader identified exactly this gap, and I see no reason to disagree: the paper's Eqs. (1) and (2) are formally incomplete because the MDP components P(s'|s,a), A(s), and the method for solving the Bellman optimality equation in continuous state space are absent. I do not find an internal contradiction in the experimental narrative; the simulation and robot demonstrations are plausible, but the missing definition is a genuine correctness and reproducibility risk rather than a disagreement with community consensus. Secondary concerns, such as the anomaly threshold being fitted to the training envelope and the absence of error bars, are real but less load-bearing; they would be largely addressed by the same request for implementation details and a fuller evaluation. Because the identified concern matches the reader's weakest assumption and supports the conditional verdict rather than overturning it, the recommendation is unchanged.","tokens_in":34263,"tokens_out":5704,"duration_ms":61680,"concrete_test":"Ask the authors to specify the exact transition model and action discretization used to compute Eq. (2) for Tables 1 and 2, or to release the segmentation code. Then rerun the simulated box-pushing experiment (Sec. 7.1) under two plausible deterministic models, e.g., s' = s + a*dt and s' = s + a*dt with a minimum-distance obstacle-avoidance term, keeping all other hyperparameters fixed. If the BNG-IRL segmentation and Table 1 metrics change materially, the reported results depend on the unspecified Q-value model; if they are invariant, the method is robust to the missing detail and the paper only needs a clarifying paragraph.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Algorithm 1 line 4 and Eqs. (1)-(4). BNG-IRL's segmentation likelihood p(s_i,a_i|g) depends on the optimality score epsilon_d = Q_Td / V*. Computing V* via Eq. (2) requires a transition model P(s'|s,a) and an action set A(s), but neither is defined anywhere. Section 4.1 states only that S is 'all reachable robot EEF poses' and actions are 'EEF velocities'; the state space does not include object or box configuration, and contact dynamics are not modeled. In a continuous 6D pose space with continuous velocities, the max over A(s) in Eq. (2) is ill-posed without a discretization or dynamics. Algorithm 1 line 4 says 'Precompute Q(·,·,Rg)' without specifying how Q or V* are obtained. Consequently epsilon_d, and hence the segmentation likelihood in Eq. (4), is not computable from the paper as written. This is not an implementation detail: the numerical segmentation results in Tables 1 and 2 depend on the hidden choice of transition model and action discretization. If the authors used a simple kinematic s' = s + a*dt model, it is inconsistent with the contact phases such as pushing or locking, where the next EEF pose is not a deterministic function of s and a alone. If they used a learned or empirical model, it must be described. Without this, the central segmentation contribution cannot be independently reproduced or checked.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an incremental learning-from-demonstration framework for contact-rich manipulation. User demonstrations are segmented by a new unsupervised algorithm, BNG-IRL, which combines feature-space Gaussian clustering with subgoal-based inverse reinforcement learning; each inferred skill receives a subgoal region, a feature constraint region, and an assigned set of demonstrations. The low-level execution model uses GMM/GMR to generate pose-conditioned velocity and force commands, and the anomaly detector computes Mahalanobis distances against the trained skill model, with a two-step confidence scheme intended to separate epistemic and aleatoric uncertainty. Skills are organized in a task graph, and recovery behaviors are incrementally appended when novel anomalies are detected. The experiments evaluate segmentation on a simulated box-pushing task and on two real robots (DLR LWR IV and DLR SARA), including a box grasping and locking task with four injected anomaly types, and compare anomaly detection against vision-language and supervised video baselines.","tokens_in":34629,"tokens_out":6442,"duration_ms":64180,"significance":"If the central segmentation derivation can be made concrete, this would be a valuable applied contribution: the framework addresses a real gap by learning hierarchical, force-sensitive manipulation skills from very few demonstrations on physical robots, and it explicitly tackles incremental recovery. The real-robot experiments, the comparisons against several baselines, and the honest limitations section in Sec. 7.4 are strengths. The segmentation idea of combining subgoal intentions with feature constraints is sensible, and the paper is right that the segmentation itself is not circular because BNG-IRL infers skills from the data rather than assuming them. However, the key algorithmic step, the computation of Q-values and the optimal value function in Eqs. (1) and (2), is never specified, which currently prevents independent reproduction and leaves the numerical segmentation results in Tables 1 and 2 dependent on an undisclosed transition model and action discretization.","major_comments":[{"comment":"The optimality score ε_d(s_i,a_i,g) that drives the segmentation likelihood (4) is not computable as stated. Equations (1) and (2) define Q_Td and V* through Bellman recurrences over a transition model P(s′|s,a) and an action set A(s), but neither is specified anywhere. Section 4.1 says only that S is 'all reachable robot EEF poses' and actions are 'EEF velocities'; for the contact tasks S is 6D with quaternion orientation (Sec. 7.3.2). In this continuous setting the max over A(s) in Eq. (2) is undefined without a discretization or a dynamical model, and in contact phases such as pushing or locking a simple kinematic s′ = s + a·dt model is inconsistent because the next EEF pose depends on the object and the contact state. If a learned or empirical transition model was used, it must be described; if a kinematic model was used, its validity in the presence of contact must be justified. Algorithm 1 line 4 ('Precompute Q(·,·,R_g)') does not specify how Q and V* are obtained. This is load-bearing because the numerical segmentation results in Tables 1 and 2 depend on the hidden choice, and the same issue affects the BN-IRL baseline reimplementation. Please provide the exact transition model, action discretization, discount factor, and reward used in every experiment.","section":"§5.1.1, Eqs. (1)–(4), Algorithm 1 line 4"},{"comment":"The anomaly detection thresholds are defined as training-data extrema: D_M,max = max over the training Mahalanobis distances and P_min(s) = min over the training set of P(s). This makes the detector a comparator against the most extreme training observation rather than a calibrated predictor, and the 'confidence' threshold P_min(s) is not a principled uncertainty level. Consequently, the paper's claim that the method leverages epistemic and aleatoric uncertainty to make confident predictions (Secs. 6.1 and 8) overstates what these quantities can certify. The empirical results may still be valid, but the authors should either provide a calibration or ablative analysis showing how the thresholds behave with varying amounts of training data, or revise the uncertainty/confidence language so that it matches the operational definition used in Eqs. (16) and (17).","section":"§6.1, Eqs. (15)–(17)"},{"comment":"The conclusion that the method is 'the only one to confidently detect all anomalies, while not triggering any false positive detection during successful executions' is only partially supported by the reported tables. For anomaly IV ('box missed'), Table 4 shows GMR(our) with recall 34.6 and a detection delay of 4.3 s, and the text states that ConditionNET performs better in this case. Table 5 reports 100% task-level accuracy for the same case, so the two tables are not necessarily contradictory, but the phrase 'confidently detect all anomalies' obscures the fact that frame-wise detection of case IV is weak and late. Please state explicitly that detection for case IV usually occurs at the execution level and after a long delay, and report the successful-execution false-positive result alongside the per-case detection accuracy so that the comparison is not read as frame-wise superiority on every anomaly.","section":"§7.3.6, Tables 4 and 5"}],"minor_comments":[{"comment":"The heading contains a typo: 'Probabilistc Feature Clustering' should be 'Probabilistic Feature Clustering'.","section":"§5.1.2 (heading)"},{"comment":"The pseudocode is ambiguous about the loop over subgoal candidates: line 3 sets 'g←s_i' inside a loop over i, but it is not clear whether Q is precomputed once per candidate state or once per demonstration, and how the precomputed Q-values are indexed during the Gibbs updates. Please rewrite the pseudocode to make the subgoal candidate indexing and the Q-value precomputation explicit.","section":"Algorithm 1, lines 3–4"},{"comment":"After the deterministic-transition simplification, the notation R_g(s,a) appears in the last line of Eq. (1), while the reward was defined as R_g(s,a,s′) = 1(s′=g). Please define the reduced reward R_g(s,a) or write the expectation over s′ explicitly to avoid ambiguity.","section":"Eq. (1), §5.1.1"},{"comment":"For the GMRwo rows in the 'slide locked' and 'box missed' columns, precision, recall, and F1 are all 0 but a numeric detection delay is still reported (8.3 and 6.6). A delay for a detector that never fires is not meaningful; mark these entries as 'no detection' or define the delay for missed detections.","section":"Table 4"},{"comment":"The anomaly trigger condition is described as 'more than 300 ms' in the text (Secs. 7.2.2 and 7.3.6), but the parameter ε in Eq. (16) and Algorithm 2 is not given in cycles or seconds. Please specify the control rate and the value of ε used in the experiments.","section":"§7.3.6"},{"comment":"The paper would benefit from a supplementary release of the code or a detailed appendix containing the discretization and dynamics used for the Q-value computation; without this, even after adding a textual description, the segmentation results in Tables 1 and 2 cannot be independently checked.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and the applied results look plausible, but the central algorithmic step (Q-value computation for BNG-IRL) is underspecified in a way that currently prevents reproduction. I do not see a circularity problem in the segmentation; the issue is missing model details rather than flawed reasoning. If the authors can supply the exact transition model, action discretization, and reward specification, and can temper or support the uncertainty-quantification claims with additional analysis, the paper would likely be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on learning from demonstration for contact-rich manipulation: this paper has a genuinely useful idea and real-robot experiments, but the central algorithm has a load-bearing gap that needs to be fixed before the results can be trusted. The new combination is BNG-IRL: subgoal-driven inverse RL fused with probabilistic feature clustering in a Bayesian nonparametric mixture, plus an extended-horizon optimality score that looks at the whole trajectory toward a subgoal rather than just the immediate action. That is a real step beyond prior BN-IRL and CBN-IRL. The two-step anomaly detection—epistemic confidence gate followed by aleatoric Mahalanobis distance—is a sensible, practical contribution, and the comparison against supervised vision-language baselines on force-based anomalies is a useful data point. Credit where due: they show a contact-rich box-locking task learned from three demonstrations, and their detector catches all four anomalies in their test set with no false positives on the successful runs. The soft spots are the problem. The stress-test note is right: Eqs. (1)–(2) require a transition model P(s'|s,a) and an action set A(s), and neither is ever specified. The state space is \"all reachable EEF poses\" and actions are \"EEF velocities,\" but the Bellman equations sum over s' and maximize over a without any dynamics or discretization. In continuous 6D pose space, that max is ill-posed. This is not a minor implementation detail—Algorithm 1 precomputes Q-values and the segmentation likelihood in Eq. (4) depends directly on them, so the numerical segmentation results in Tables 1–2 cannot be independently reproduced or even checked. The anomaly thresholds are also fitted to the training data envelope (max Mahalanobis distance, min density), which is okay as a heuristic but should be described as such; the claim about \"confidently detecting all anomalies\" is really \"detecting deviations beyond the tightest training variation,\" and the \"box missed\" case shows the method's limit. No code, no data, no error bars; several claims about reduced compute and data efficiency are not backed by numbers. In proportion: the experimental story is plausible, and the framework ideas are sound enough to deserve a proper referee. But as submitted, the segmentation core is not computable as written. This is a serious revise-and-resubmit. If they clarify the dynamics (even a simple kinematic model with explicit discretization), release code, and re-frame the anomaly-threshold fitting honestly, it could be a solid contribution. I'd send it to review, not desk-reject, but I would insist on those fixes.","headline":"A promising LfD framework with real-robot results, but the segmentation algorithm's Q-value computation is never defined, which makes the paper's core contribution non-reproducible as written.","tokens_in":35145,"tokens_out":1516,"would_cite":false,"duration_ms":17494,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T40","68T05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A robot can learn a multi-step contact task from three demonstrations and recover from never-seen failures.","keywords":["learning from demonstration","unsupervised task segmentation","inverse reinforcement learning","anomaly detection","incremental learning","contact-rich manipulation","task graph","Gaussian mixture regression"],"falsifier":"Recompute the box-pushing segmentation from the same demonstrations using a different deterministic transition model, such as a different velocity discretization or a small perturbation, and check whether the inferred skill boundaries and subgoals stay the same; if they change, the claimed intention recognition is sensitive to the unstated dynamics rather than being determined by the demonstrations.","tokens_in":34066,"feed_emoji":"🤖","tokens_out":10084,"duration_ms":91357,"temperature":0.7,"pith_summary":"This paper claims that a robot can learn a hierarchical model of a multi-step contact-rich manipulation task from as few as three human demonstrations on the real system, without simulation, without skill labels, and without a list of possible failures. The learned model has two levels: each skill is a low-level motion and force model with an expected feature region and a subgoal, and the skills are organized in a task graph that monitors progress and selects recovery behaviors. The authors argue that segmenting the demonstrations with a combination of inverse reinforcement learning and probabilistic feature clustering, rather than clustering trajectories or forces alone, is what makes the skill boundaries meaningful. They further claim that conditioning anomaly detection on the measured end-effector pose rather than on time lets the robot tell apart unfamiliar situations from normal variability, and react accordingly. If the claims hold, contact tasks that are expensive to simulate can be taught quickly, monitored online, and extended with recovery behaviors only when genuinely new failures occur.","feed_headline":"Three demos teach a robot a contact task and its recovery","feed_subtitle":"Unsupervised skill segmentation and force-based monitoring let a robot spot and recover from never-seen failures.","key_machinery":"The load-bearing object is the optimality score $\\varepsilon_d(s_i,a_i,g_k^d)=Q^{T_d}(s_i,a_i,g_k^d)/V^*(s_i,g_k^d)$, the ratio between the action value of the demonstrated policy and the optimal value for reaching a candidate subgoal; it lies in $[0,1]$, is fed through a softmax to give the probability that the observed action is directed to a given subgoal, and is multiplied by a multivariate Gaussian feature-constraint likelihood in the observation model (4). This combined likelihood is the basis for a Bayesian nonparametric Gaussian mixture with a Chinese restaurant process prior and Normal-Inverse-Wishart priors, inferred by collapsed Gibbs sampling (Algorithm 1). On the execution side, the machinery is Gaussian mixture regression conditioned on the measured end-effector pose: the predicted mean velocity and force and the conditional covariance define a Mahalanobis distance (15) whose threshold $D_{M,\\max}$ is the largest distance over training data, while the mixture density (14) controls whether the prediction is confident. Together these pieces let the framework segment demonstrations unsupervised, monitor execution online, and decide when to learn a new recovery behavior.","core_discovery":"The central claim is that a task performed in contact with the environment can be decomposed into skills by jointly scoring how each demonstrated action is directed toward a reachable subgoal and how the accompanying multimodal features cluster in a shared probabilistic model. The resulting algorithm, Bayesian Nonparametric Gaussian Inverse Reinforcement Learning (BNG-IRL), models each skill as a mixture component with a subgoal region in end-effector state space and a constraint region in feature space, and infers the number of skills, the skill assignments, and the subgoal of every demonstration by collapsed Gibbs sampling. On the execution side, the framework predicts the expected end-effector velocity and force from the measured pose via Gaussian mixture regression, flags a deviation when the Mahalanobis distance from the prediction exceeds the largest distance seen in training, and only trusts the flag when the measured pose lies in a region of low epistemic uncertainty. If a novel anomaly is confirmed, the user demonstrates a recovery behavior that is segmented and appended to the task graph at the skill where the anomaly occurred. The paper reports that this pipeline learns the box grasping and locking task from three demonstrations, detects four different anomaly types on a real robot without false positives during successful runs, and outperforms the compared segmentation and anomaly-detection baselines.","pith_inferences":["A direct stress test the paper leaves implicit is to alter the hidden transition model while keeping the demonstrations fixed; stable skill boundaries would indicate the segmentation is driven by the demonstrations' intentions, not by the model's assumptions.","The known weak case, the visually obvious but force-neutral 'box missed' anomaly, suggests a parallel fusion of this pose-conditioned force detector with a vision-based skill-state monitor, with the task graph arbitrating between their reports.","Because each skill's feature region is a multivariate Gaussian, the method can only capture linear correlations between features; a kernelized or mixture extension would be a natural test of whether nonlinear force-position trade-offs can still be learned from the same few demonstrations.","The paper considers reusing recovery skills across different skills; if skill preconditions were learned instead of hard-coded by graph position, one demonstrated recovery could in principle serve several skills, which would cut the number of required recovery demonstrations."],"forward_implications":["Contact-rich tasks that are hard to simulate can be taught on the real robot from three kinesthetic demonstrations, without an accurate physics model.","Skill boundaries found by combining subgoal intentions with feature constraints are more reliable than clustering on trajectories or forces alone, as measured against four segmentation baselines.","A detector that conditions on the measured end-effector pose can flag force deviations before they become visually observable, and it will not raise alarms in regions of the state space it does not know yet.","When a new anomaly is confirmed, the demonstrated recovery is added to the task graph for that skill, so the same recovery works regardless of when during the skill the anomaly appears.","Distinguishing aleatoric from epistemic uncertainty makes the anomaly detector more sensitive where training variance is low and switches to user supervision where the model has no data."],"supporting_citations":[{"why":"It supplies the Bayesian nonparametric inverse reinforcement learning formulation and the sparse-subgoal reward idea that BNG-IRL extends with feature clustering and a longer evaluation horizon.","marker":"Michini et al. (2015)"},{"why":"It provides the constraint-based BN-IRL approach whose computational bottleneck motivates separating feature-constraint inference from Q-value computation.","marker":"Park et al. (2020)"},{"why":"It provides the infinite Gaussian mixture model used as the clustering backbone and as the BNGMM baseline.","marker":"Rasmussen (1999)"},{"why":"It supplies the physically consistent BN-GMM baseline whose velocity-similarity clustering BNG-IRL is compared against in simulation.","marker":"Figueroa and Billard (2018)"},{"why":"It supplies the Bayesian online changepoint segmentation baseline evaluated on both the simulation task and the box-grasping task.","marker":"Sugawara et al. (2023)"},{"why":"It supplies the automatic waypoint extraction baseline used for trajectory segmentation in both experimental settings.","marker":"Shi et al. (2023)"},{"why":"It supplies ConditionNET, the supervised vision-language anomaly-detection baseline that the unsupervised detector is compared against.","marker":"Sliwowski and Lee (2024)"},{"why":"It supplies FinoNET, the supervised deep failure-detection baseline used in the anomaly detection comparison.","marker":"Inceoglu et al. (2021)"},{"why":"It supplies the vision-language chain-of-thought question-answering baseline for online anomaly detection.","marker":"Agia et al. (2024)"}],"fun_headline_variants":["Three demos teach contact task, skills, and unseen-failure recovery","Robot learns force-based contacts and error recovery from 3 demos","Unsupervised skill split enables touch-based error detection and repair","Few demos, force monitoring, and auto-recovery for contact tasks","BNG-IRL: nonparametric task decomposition for robot error recovery"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The segmentation relies on an assumed but never specified model of how robot states change under actions and how the optimal value function is computed; if that transition model and action discretization are wrong, the optimality scores that drive skill inference lose their meaning.","fun_headline_variants_meta":{"raw":{"variants":["Three demos teach contact task, skills, and unseen-failure recovery","Robot learns force-based contacts and error recovery from 3 demos","Unsupervised skill split enables touch-based error detection and repair","Few demos, force monitoring, and auto-recovery for contact tasks","BNG-IRL: nonparametric task decomposition for robot error recovery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00114,"raw_usage":{"total_tokens":4757,"prompt_tokens":998,"completion_tokens":3759,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":3667}},"tokens_in":614,"tokens_out":3759,"duration_ms":27047,"temperature":1.0,"reasoning_tokens":3667,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:25:48.484128+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the box-pushing segmentation from the same demonstrations using a different deterministic transition model, such as a different velocity discretization or a small perturbation, and check whether the inferred skill boundaries and subgoals stay the same; if they change, the claimed intention recognition is sensitive to the unstated dynamics rather than being determined by the demonstrations.","supporting_citations":[{"cited_title":"In: Conference on Robot Learning","cited_arxiv_id":null,"evidence_quote":"It provides the constraint-based BN-IRL approach whose computational bottleneck motivates separating feature-constraint inference from Q-value computation."}],"review_version":1}