{"id":"2b01cb5f-1d2b-4318-b72e-86120c51f58a","arxiv_id":"1908.07263","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Force-guided Bayesian optimization locally updates the position and duration means of one hidden semi-Markov model state, enabling spatial and temporal trajectory deformation in about five evaluations.","lead":"A robot learns a task from demonstrations and then can reshape its path when a human pushes it, by treating the push as a signal. The method uses Bayesian optimization to adjust only the local part of the learned motion, so adaptation takes only a handful of trials.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported convergence is measured against a scripted external force, not a human in the loop; under real closed-loop interaction the force objective is non-stationary and underdetermined, so the central claim about human-guided adaptation is not yet supported.","rationale":"The reader's weakest assumption is that interaction forces are a reliable, interpretable signal of human intent. I agree and sharpen that concern: even if force is informative at the instant of contact, the proposed objective is not a stable function of θ in the interactive loop. The synthetic force profile in Section IV removes the human response, so the reported convergence in Section IV-C is an artifact of an exogenously fixed disturbance. This makes the central claim about human-guided adaptation unverified, rather than mathematically invalid. A closed-loop test with a known simulated human would settle whether the objective is identifiable. Since the gap is empirical rather than a demonstrated inconsistency of the formalism, the appropriate verdict remains conditional pending that test; no change to the reader's verdict is needed.","tokens_in":12123,"tokens_out":6488,"duration_ms":80118,"concrete_test":"Run a closed-loop simulation with a simulated human whose target trajectory is known: the human is modeled as an impedance controller pulling the robot toward a target deformation, and the robot's reference is generated from the nominal HSMM and updated by the proposed force-guided BayesOpt of Section III-B. The force signal f_s is produced dynamically by the human-robot interaction, not prescribed as in Section IV. After convergence, compare the recovered θ* to the ground-truth target state mean and duration, and also enumerate all θ in the 2σ search box that yield ||\\hat f - f_s|| below sensor noise. If the minimizer is non-unique or does not match the target, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Abstract, Section III-B) is that a robot can 'quickly adapt' learned trajectory distributions from human-applied forces by minimizing disagreement forces. The load-bearing step is the definition of the objective. In Section III-B the observations are D_n = {(θ_i,j, ||\\hat f - f_s||_j)}, so the scalar quantity being minimized is an L2 force error. This is a well-posed objective only if the force signal is an exogenous, repeatable function of θ. That is exactly the setup in Section IV: a 'simulated external force' with a fixed direction and profile is applied to the end-effector, independent of the human's response. In a real physical interaction the force is generated by the coupled human-robot system: it depends on the human's intended trajectory and on the robot's current reference, and it vanishes as soon as the robot tracks the human. Consequently (i) the same θ can produce different force readings depending on human reaction, violating the GP stationarity assumption in Section III-A, and (ii) many deformed trajectories are force-free, so minimizing the norm cannot identify which of them encodes the human's intent. The paper does not test the interactive case; its future-work section states that real scenarios are left for later work. Thus the evidence presented does not support the claim that the adapted distribution is 'consistent with ... the changes introduced by the human'.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for adapting a demonstration-learned trajectory distribution through physical human-robot interaction. A hidden semi-Markov model (HSMM) encodes the nominal task plan, and a smooth reference trajectory distribution is retrieved by weighted least-squares over static and dynamic features. During an adaptation phase, the robot minimizes disagreement forces by running Bayesian optimization locally on the mean motor-command and duration parameters of the currently active HSMM state, with the search domain extracted from the model covariances. Experiments on synthetic 2D C- and L-shape pick-and-place tasks, using a simulated external force with a fixed direction and profile, show convergence in five to six BayesOpt iterations and local spatial and temporal deformations of the reference trajectory. The paper claims that this yields data-efficient adaptation consistent with human-intended changes, but real human-in-the-loop experiments are explicitly left to future work.","tokens_in":12397,"tokens_out":4093,"duration_ms":47070,"significance":"If the central claim holds, the paper makes a useful contribution to learning-from-demonstration and human-robot collaboration: it combines a probabilistic task model with force-guided Bayesian optimization in a way that preserves the nominal plan while incorporating local corrections. The HSMM forward computation and the weighted least-squares trajectory retrieval are standard and are presented correctly. The local BayesOpt formulation with expected improvement is also standard, and the use of the active HSMM state to reduce dimensionality is a sensible and clearly explained choice. These strengths make the framework plausible and worth developing. However, the empirical evidence is limited to a single scripted force profile in simulation, so the paper's central claim about adaptation to human-intended changes is not yet supported.","major_comments":[{"comment":"The only experimental instantiation of the human is a simulated external force with a fixed direction and profile (Fig. 4), and Section VI explicitly states that real scenarios are left for future work. In the objective D_n = {(θ_i,j, ||f_hat - f_s||_j)} defined in Section III-B, the GP stationarity assumption of Section III-A requires the force to be an exogenous, repeatable function of θ. In a closed-loop physical interaction, the force is generated by the coupled human-robot system and depends on the human's reaction, so the same θ can produce different force readings. The presented evidence therefore does not support the abstract's claim that the adapted distribution is consistent with the changes introduced by the human. A concrete remedy is a human-subjects study with time-varying intentions, or a substantially softened claim that restricts the result to open-loop force corrections.","section":"Section IV-C / Section III-B"},{"comment":"The force-minimization objective is underdetermined as a model of human intent. As the robot deforms its trajectory toward the human's preferred motion, the interaction force tends to zero, so many different deformed trajectories are equally force-free. Minimizing ||f_hat - f_s|| cannot by itself identify which of these force-free trajectories encodes the intended change. The paper assumes in the opening of Section III that force cues provide information about an unobservable reward function, but it neither formalizes nor tests the needed identifiability or monotonicity conditions. Please provide a formal statement of the assumptions under which the force error uniquely identifies the intended adaptation, and add experiments with multiple force directions and profiles that correspond to distinct intended changes.","section":"Section III-B"},{"comment":"The comparison with full-parameter BayesOpt over θ ∈ R^18 is reported as not converging, and this comparison underpins the claimed benefit of local search. The paragraph does not state the number of iterations, the GP kernel or hyperparameters, the acquisition-function settings, or the initialization for this baseline. Without these details the reader cannot judge whether the failure is due to dimensionality alone or to an unfair or poorly tuned baseline. Please report a controlled comparison.","section":"Section IV-C"},{"comment":"The time window T_w and the re-computation of the state sequence s_{t:T_w} are central to the online adaptation behavior, but the paper does not specify how T_w was chosen or how sensitive the results are to this choice and to the search bounds (±2σ_i). Since these quantities define the local search space, please report a sensitivity analysis or at least a justification for the chosen values.","section":"Section III-C"}],"minor_comments":[{"comment":"The parameter d_max in Eq. (1) is used before being defined; please state its meaning and how it is selected.","section":"Section II-A"},{"comment":"In Eq. (17), the threshold τ is not defined precisely; please state whether it is the incumbent value or a user-specified target, and how it is updated during the optimization.","section":"Section III-A"},{"comment":"The text says six synthetic demonstrations were collected, but no details are given about the generation process, sensor noise, or number of independent runs; please add these experimental details for reproducibility.","section":"Section IV-A"},{"comment":"Figure 5 shows GP posterior means and credible intervals, but the locations of the queried points and the order of evaluations are not shown; adding them would make the reported convergence to six and five iterations more transparent.","section":"Section IV-C / Fig. 5"},{"comment":"There are several typos, including 'leaded' in Section IV-C and 'forge-guided' in Section IV-A; please proofread the manuscript.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a well-written proof-of-concept with a sound mathematical core, but the gap between the claimed human-interactive adaptation and the scripted-force synthetic evaluation is substantial. If the authors can add human-in-the-loop experiments or clearly reframe the contribution as an open-loop force-correction mechanism with an explicit identifiability analysis, the paper could become acceptable. I do not see an internal inconsistency that would force rejection, but the central claim currently outruns the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Leonel Rozo's paper combines learning-from-demonstration with Bayesian optimization for physically interactive trajectory adaptation. The idea is to learn a nominal task plan as an HSMM, then, when a human applies force, use BayesOpt to locally adjust the current state's position and duration means to minimize the sensed disagreement force. The integration is genuinely new: unlike Losey and O'Malley's analytical deformation, this uses a probabilistic model and Bayesian optimization, and it handles temporal adaptation via duration means. The math is standard and correctly applied: the HSMM forward computation and the weighted least-squares trajectory retrieval are properly derived. Restricting the search space to the local state distribution is a clean way to reduce dimensionality while respecting the learned structure. The synthetic experiments demonstrate fast convergence (5-6 iterations) and localized adaptation, leaving the start and end of the trajectory intact.\n\nThe main weakness is the evaluation. The experiments use a single simulated force profile, no physical robot, no human subjects, no baselines, and no quantitative error bars—just two convergence plots. The central assumption that interaction forces are a reliable, monotonic signal of the human's intended adaptation is stated but never tested. More importantly, the simulated force is exogenous and independent of the robot's response; in real human-robot interaction, the force arises from the coupled system and will vanish as the robot accommodates. That makes the objective non-stationary and underdetermined: many deformed trajectories are force-free, so minimizing the force norm cannot uniquely identify the human's intent. The stress-test note makes this point well, and the paper's future-work section concedes that real scenarios are left for later. So the claim about consistency with 'the changes introduced by the human' is not yet supported by the evidence.\n\nWho is this for? Someone working in haptic interaction or learning-from-demonstration might find the integration useful as a building block. But as a standalone paper, it needs substantial additional evidence. I would send it to peer review because the idea is plausible and the formulation is clean, but I would expect the referee to demand real robot experiments with human subjects, baselines (e.g., against Losey and O'Malley and against global BayesOpt), and quantitative metrics with variance. If this is for a workshop, it's fine as is; for a full conference, it needs the experiments.","headline":"A clean integration of local Bayesian optimization with HSMM trajectory distributions that adapts spatial and temporal patterns from force cues, but the evidence is too thin to support the human-interaction claims.","tokens_in":12922,"tokens_out":3857,"would_cite":false,"duration_ms":34004,"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 physical forces applied by a human can guide a robot to locally adapt its demonstration-learned trajectory through Bayesian optimization, converging in about five to six trials.","keywords":["Bayesian optimization","hidden semi-Markov model","learning from demonstration","physical human-robot interaction","trajectory adaptation","force-guided adaptation","haptic communication","Gaussian process"],"falsifier":"Run the same framework with human subjects pushing the end-effector in a direction that increases rather than decreases the force norm (for example, pressing against the robot's preferred path to indicate a detour). If the adapted trajectory does not follow the human's indicated direction, or if the disagreement force is not driven to near zero because the human's intent is not encoded as force minimization, the central claim fails. A real-human dataset with noisy, non-monotonic force readings relative to the intended target would settle the assumption.","tokens_in":11914,"feed_emoji":"🤖","tokens_out":4687,"duration_ms":42784,"temperature":0.7,"pith_summary":"The paper claims that a robot can adapt a task trajectory learned from demonstrations by letting a human physically push it, and treating the sensed interaction forces as evidence about what the human wants changed. The proposed method uses Bayesian optimization to adjust only the locally relevant parameters of a hidden semi-Markov model that encodes the nominal task plan, so the search stays low-dimensional and data-efficient. A simulated pick-and-place experiment shows convergence in five to six force-guided iterations, producing deformed trajectory distributions that match the human-imposed change while preserving the parts of the plan the human did not touch. The point of the work is to give robots a way to incorporate human haptic corrections without handcrafting a reward or objective function.","feed_headline":"Six force-guided trials adapt a learned robot trajectory","feed_subtitle":"Local Bayesian optimization reshapes only the part of the plan the human pushes, leaving the rest intact.","key_machinery":"The central object is a hidden semi-Markov model (HSMM), a probabilistic encoding of demonstrated trajectories that includes explicit lognormal state-duration distributions alongside per-state Gaussian observation models. It carries the argument in two roles: it supplies the nominal trajectory distribution via a smooth trajectory-retrieval step, and it identifies which state should be locally adapted when a force cue arrives, which restricts Bayesian optimization to a low-dimensional parameter vector. The adaptation itself is driven by expected-improvement Bayesian optimization with a Gaussian process surrogate, minimizing the norm of the difference between the reference and sensed forces.","core_discovery":"The central claim is that interaction forces, interpreted as noisy cues to an unobservable objective the human is trying to optimize, can drive a local, data-efficient adaptation of a learned trajectory model. Concretely, the paper proposes to minimize the disagreement between the reference force distribution and sensed forces by running Bayesian optimization over the mean of the motor commands and the mean duration of the HSMM state in which the adaptation is triggered. Because the active state is identified from the HSMM forward variable, the parameter vector remains tiny, and the search domain is automatically bounded by the learned state covariance. The paper reports that this local search converges in five or six iterations and yields reference trajectory distributions that deform only in the adapted region, whereas optimizing all HSMM means at once fails due to the curse of dimensionality.","pith_inferences":["A natural next step, which the paper does not test, is to run the method with real human force data rather than a synthetic force profile; the assumption that force magnitude is monotonically related to disagreement would then face a harder test.","The local-state selection via the forward variable could be extended to simultaneous adaptation of several neighboring states, since a human push often spans more than one HSMM state.","If the force-to-intent assumption holds, the same mechanism could serve as a negotiation protocol in shared tasks: the robot resists deviation from the nominal plan unless the human's physical input is consistent enough to shift the model.","Adapting covariance parameters is left open; methods that respect positive-definiteness, such as optimizing in a transformed space, would let the robot also change the variance of the trajectory distribution."],"forward_implications":["If the method is correct, a robot can incorporate a human's physical correction into a demonstration-learned skill without retraining the model from scratch.","Because the adaptation is local, the trajectory returns to the nominal plan when the human stops pushing, which suits tasks with fixed start and end constraints.","Since the search domain is inferred from the model's own covariance, the method avoids an explicit handcrafted reward function and needs only a handful of physical trials.","Spatial and temporal patterns can be adapted together, because the optimized parameters include both the state mean and the state duration.","The Bayesian optimization backbone leaves room to add constraints, such as limiting deviation from the nominal plan or penalizing jerky motion."],"supporting_citations":[{"why":"Supplies the Bayesian optimization framework, including expected improvement and Gaussian process surrogates, that the adaptation loop is built on.","marker":"[7]"},{"why":"Demonstrates BayesOpt for choosing collaborative actions in physical human-robot interaction, the line of work this paper extends by starting from a demonstration-learned policy.","marker":"[12]"},{"why":"Uses BayesOpt to learn an unknown reward for a physical handover skill, another prior HRC application that lacks the demonstration-based nominal model.","marker":"[13]"},{"why":"Defines hidden semi-Markov models, the encoding used for the nominal task plan.","marker":"[14]"},{"why":"Provides the HSMM-based trajectory generation with static and dynamic features that turns the model into a reference trajectory distribution.","marker":"[15]"},{"why":"Presents the analytical trajectory-deformation baseline that this method contrasts with, since it uses gradient-based optimization instead of BayesOpt and only spatial deformations.","marker":"[16]"},{"why":"Motivates defining a safe or restricted search domain for BayesOpt, analogous to deriving the local bounds from the learned state covariances.","marker":"[17]"},{"why":"Gives the HMM forward-variable recursion that underlies the HSMM state-sequence retrieval used to identify the active state.","marker":"[22]"},{"why":"Supplies the efficient forward-backward implementation for explicit-duration HSMMs used to compute the forward variable.","marker":"[23]"},{"why":"Supports the reference-point-dependent probabilistic model approach to generating motions from learned distributions.","marker":"[24]"}],"fun_headline_variants":["Five force trials locally adapt a learned trajectory","Force-guided Bayesian tweaks reshape only the pushed part","Five force-guided trials deform one segment of a plan","Local Bayesian opt from human force adapts a robot path"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the force a human applies is a trustworthy, informative signal of the intended trajectory change, so that minimizing the measured force disagreement is the same as fulfilling the human's intent; the paper assumes this rather than testing it with real human forces.","fun_headline_variants_meta":{"raw":{"variants":["Five force trials locally adapt a learned trajectory","Force-guided Bayesian tweaks reshape only the pushed part","Five force-guided trials deform one segment of a plan","Local Bayesian opt from human force adapts a robot path"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000508,"raw_usage":{"total_tokens":2450,"prompt_tokens":894,"completion_tokens":1556,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":1506}},"tokens_in":510,"tokens_out":1556,"duration_ms":10873,"temperature":1.0,"reasoning_tokens":1506,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:20:53.626661+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same framework with human subjects pushing the end-effector in a direction that increases rather than decreases the force norm (for example, pressing against the robot's preferred path to indicate a detour). If the adapted trajectory does not follow the human's indicated direction, or if the disagreement force is not driven to near zero because the human's intent is not encoded as force minimization, the central claim fails. A real-human dataset with noisy, non-monotonic force readings relative to the intended target would settle the assumption.","supporting_citations":[{"cited_title":"Taking the human out of the loop: A review of Bayesian optimiza- tion,","cited_arxiv_id":null,"evidence_quote":"Supplies the Bayesian optimization framework, including expected improvement and Gaussian process surrogates, that the adaptation loop is built on."},{"cited_title":"A sensorimotor reinforcement learning framework for physical human-robot interaction,","cited_arxiv_id":null,"evidence_quote":"Demonstrates BayesOpt for choosing collaborative actions in physical human-robot interaction, the line of work this paper extends by starting from a demonstration-learned policy."},{"cited_title":"Learning dynamic robot-to-human object handover from human feedback,","cited_arxiv_id":null,"evidence_quote":"Uses BayesOpt to learn an unknown reward for a physical handover skill, another prior HRC application that lacks the demonstration-based nominal model."},{"cited_title":"Hidden semi-Markov models,","cited_arxiv_id":null,"evidence_quote":"Defines hidden semi-Markov models, the encoding used for the nominal task plan."},{"cited_title":"Learning controllers for reactive and proactive behaviors in human-robot col- laboration,","cited_arxiv_id":null,"evidence_quote":"Provides the HSMM-based trajectory generation with static and dynamic features that turns the model into a reference trajectory distribution."},{"cited_title":"Trajectory deformations from physical human–robot interaction,","cited_arxiv_id":null,"evidence_quote":"Presents the analytical trajectory-deformation baseline that this method contrasts with, since it uses gradient-based optimization instead of BayesOpt and only spatial deformations."},{"cited_title":"Safe controller opti- mization for quadrotors with Gaussian processes,","cited_arxiv_id":null,"evidence_quote":"Motivates defining a safe or restricted search domain for BayesOpt, analogous to deriving the local bounds from the learned state covariances."},{"cited_title":"A tutorial on hidden Markov models and selected appli- cations in speech recognition,","cited_arxiv_id":null,"evidence_quote":"Gives the HMM forward-variable recursion that underlies the HSMM state-sequence retrieval used to identify the active state."},{"cited_title":"Practical implementation of an efficient forward-backward algorithm for an explicit-duration hidden Markov model,","cited_arxiv_id":null,"evidence_quote":"Supplies the efficient forward-backward implementation for explicit-duration HSMMs used to compute the forward variable."},{"cited_title":"Learning, generation, and recognition of motions by reference-point-dependent probabilistic models,","cited_arxiv_id":null,"evidence_quote":"Supports the reference-point-dependent probabilistic model approach to generating motions from learned distributions."}],"review_version":1}