{"id":"ad68b6f4-7b87-4882-bd23-8188db47b597","arxiv_id":"2412.15756","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A neural-network latent-state friction model, trained by EM with particle filtering, predicts KUKA KR6 R700 open-loop dynamics more accurately than six baselines on a single validation trajectory.","lead":"The paper builds a probabilistic state-space model in which neural networks represent both the friction torque and an unobserved latent state that captures missing friction dynamics, and fits it to encoder and motor-current data from a KUKA robot using expectation-maximization and particle filtering. The model gives better open-loop position/velocity predictions than Coulomb, Stribeck, LuGre, GMS, NN, and RNN baselines on one held-out trajectory.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed superiority rests on a single validation trajectory with no error bars; the 10 s margin over LuGre is within noise, so the central claim is not yet established.","rationale":"The paper has real strengths: a held-out validation trajectory, a plausible latent-state formulation, qualitative friction characteristics that match known physics, and a fair set of baselines including physics-based, static, and recurrent models. However, the central claim is empirical and comparative, and the evidence is a single run with point estimates. The 10 s difference between LVM and LuGre (0.024 vs 0.027 MSE) is small enough that run-to-run stochasticity or a different validation trajectory could flip it; the full-trajectory advantage, while large, is a single observation and is accompanied by omitted baselines whose predictions diverged. The reader's weakest-assumption focuses on identifiability of the 2D Markov latent state. That is a legitimate modeling concern, but it is secondary: an unidentifiable model could still predict well on the tested trajectory, and the immediate threat to the headline claim is the lack of statistical evidence that the advantage generalizes. The concrete check of multiple validation trajectories with error bars would settle whether the LVM truly outperforms LuGre and the other baselines. Given that the paper's contribution is promising but not yet demonstrated robustly, the conditional verdict is appropriate; no change in verdict is needed.","tokens_in":9421,"tokens_out":4304,"duration_ms":39434,"concrete_test":"Run the identification and open-loop evaluation on at least K=5 independent validation trajectories (or randomly re-split the collected data multiple times) and report mean ± std (or box plots) of MSE and MAE for all models over both the 10 s and complete horizons, using identical preprocessing and hyperparameters. Also perform a latent-dimension sweep dim ∈ {1,2,3,4} and report the validation likelihood or open-loop error. If the LVM does not beat LuGre by more than the error bars on most trajectories, the 'outperforms all benchmarks' claim is not supported; if dim=2 is not selected by a visible elbow, the model is overparameterized.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (Section V-B) is that the LVM 'outperforms all benchmark models in both accuracy and robustness.' The evidence is a single held-out trajectory: Table I reports point estimates of MSE/MAE for one 31.4 s validation run. The 10 s MSE, 0.024 for the LVM vs 0.027 for LuGre, differ by ~11%, a margin that can easily be reversed by run-to-run stochasticity in EM/particle filtering, network initialization, or dataset realization. No error bars, repeated trials, or per-joint statistics are reported. For the complete trajectory the LVM's advantage is large (0.11 vs 0.66 MSE), but it is a single observation, and the Simple and RNN entries are omitted because their predictions 'exploded'—an uncontrolled outcome that makes the comparison incomplete. The paper also states (Section V-A) that the latent dimension 'was determined to be 2' without describing the selection procedure; this matters because an overparameterized latent state could absorb trajectory-specific dynamics and inflate the open-loop score without generalizing. Taken together, the empirical foundation for the headline claim is one favorable draw, not a demonstrated robust advantage.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a probabilistic state-space model for robot-joint friction identification, in which the rigid-body dynamics are augmented with a low-dimensional latent state whose transition and friction torque are parameterized by neural networks. The model parameters (inertial parameters, two NN parameter sets, and noise covariances) are estimated by maximizing the marginal likelihood via the EM algorithm with Sequential Monte Carlo approximations. The method is evaluated on a KUKA KR6 R700 by open-loop prediction of joint positions and velocities from measured motor torques, comparing against Coulomb+viscous, Stribeck, LuGre, GMS, fully connected NN, and RNN baselines. The reported results show the LVM achieving lower MSE/MAE than the baselines over a 10 s horizon and over the full 31.4 s validation trajectory.","tokens_in":9689,"tokens_out":6573,"duration_ms":58164,"significance":"If the results hold, the contribution is valuable: it offers a principled probabilistic framework for learning unobserved friction dynamics without torque sensors, integrates latent dynamics and neural friction models into the standard rigid-body model, and evaluates on a genuine held-out trajectory rather than on the training data. The full-trajectory open-loop result (MSE 0.11 vs 0.66 for LuGre) is suggestive of a real stability advantage. However, the paper's central claim of outperforming all benchmarks rests on a single validation run with no error bars and on an undocumented choice of latent dimension, so the significance is presently conditional on the empirical evaluation being strengthened.","major_comments":[{"comment":"The central claim that 'the proposed data-driven latent variable model outperforms all benchmark models in both accuracy and robustness' is supported only by point estimates in Table I from a single held-out validation trajectory. The paper reports no repeated trials, no standard errors or confidence intervals, and no per-joint breakdown; the 10 s MSE margin over LuGre (0.024 vs 0.027) is small enough that it could be reversed by EM/particle-filter randomness, network initialization, or dataset realization. The full-trajectory margin (0.11 vs 0.66) is encouraging but is still a single observation. Please add multiple validation trajectories or repeated identifications, or at minimum report variability across joints and over repeated runs, before claiming superiority.","section":"Section V-B, Table I"},{"comment":"The statement that 'the optimal dimension of the latent state z, was determined to be 2' is given without reporting the selection procedure, a sensitivity analysis, or an identifiability discussion. Because the latent state is the mechanism by which the model absorbs unmodeled friction dynamics, an overparameterized or underparameterized latent dimension could directly affect the open-loop prediction scores and generalizability. Please document how the dimension was chosen (e.g., cross-validated likelihood, held-out score, or comparison across dimensions) and report the sensitivity of Table I to the latent dimension and to the number of particles (200).","section":"Section V-A"},{"comment":"Table I omits the Simple and RNN entries for the complete trajectory because their predictions 'exploded', which makes the comparison incomplete and the robustness claim hard to quantify. Please report the actual error values even if they are very large, define a divergence criterion, and indicate at what time the divergence occurred; without this information the reader cannot distinguish a systematic robustness advantage of the LVM from a single uncontrolled divergence event.","section":"Section V-B, Table I"}],"minor_comments":[{"comment":"Since no ground-truth friction torque is available, the identified friction characteristics can only illustrate what the model learned; please reframe the text explicitly as an illustration rather than as confirmation that the model captures real friction phenomena, or add an independent quantitative check.","section":"Section V-C"},{"comment":"Please proofread carefully: there are many typographical errors, including 'Quantative' and 'en' in Table I, 'adress', 'Lugre' for 'LuGre', 'frecuency', 'camparison', 'subtilities', 'te note', 'preculding', 'analysys', 'emprical', 'incoporates', 'arbritrary', 'succesfully', 'intrinscally', and 'velocties'. The notation '9q' and '9xt' should be typeset as \\dot{q} and \\dot{x}_t throughout.","section":"Throughout"},{"comment":"The paper does not state whether code or data will be made available. For a data-driven method evaluated on a proprietary robot platform, a public release of the dataset or code, or at least a detailed experiment protocol, would substantially improve reproducibility.","section":"Section IV-A"},{"comment":"The term 'robustness' is used without a definition; please specify the metric or criterion (e.g., stability over the full trajectory, error growth rate, or worst-case error) so that the claim can be evaluated quantitatively.","section":"Section V-B"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the method is promising. The main weakness is the thin experimental validation: a single held-out trajectory with no uncertainty quantification and an undocumented latent-dimension choice. I recommend major revision with a request to strengthen the evaluation and report model-selection details. If the journal values reproducibility for data-driven papers, code/data availability should also be requested."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What's actually new: the paper combines probabilistic state-space models with neural-network friction and latent dynamics, trained via EM/SMC on noisy encoder data, and evaluates on a real KUKA. That specific integration is new compared to the LuGre/GMS, NN-residual [16], and LSTM [17] baselines. They also honestly frame the identification problem as partially unsupervised, which is a real step beyond supervised friction black boxes. The held-out evaluation is genuine: they apply measured motor torques and predict positions/velocities on a trajectory not used for training, which is the right test for a dynamics model.\n\nWhat it does well: the method is carefully positioned in the PSSM literature, the EM derivation is standard but correctly presented, and the SMC details are plausible. They state at the end that full 6-7 DOF validation is still needed, which is honest. For a KUKA without torque sensors, the open-loop prediction comparison is a reasonable proxy.\n\nWhere it's soft: the headline claim in Section V-B says the LVM 'outperforms all benchmark models in both accuracy and robustness,' but the evidence is one 31.4 s validation trajectory. No error bars, no repeated runs. Over the 10 s window, LVM MSE is 0.024 vs LuGre's 0.027: an 11% gap that could easily flip with particle-filter stochasticity or EM initialization. The complete-trajectory advantage is large (0.11 vs 0.66), but it's a single observation, and Simple and RNN are omitted after their predictions exploded, which makes the baseline table incomplete. The latent dimension of 2 is reported as 'determined' with no selection procedure; since the latent state is the mechanism claimed to capture missing dynamics, that matters. No code or data is released, so other groups can't reproduce or check the sensitivity.\n\nAlso, Section V-C's friction curve for Joint 1 is derived from the fitted model, not measured, and the paper itself says no quantitative comparison is possible. That's fine as a sanity check, but it shouldn't be read as validation.\n\nVerdict: this is a solid method proposal with a plausible but not yet demonstrated advantage. I'd send it to peer review: the framing is coherent, the evaluation protocol is appropriate in design though thin in statistical power, and the limitations are stated. I would not cite it as an established result yet, but I'd follow it for a revised version with repeated trials, latent-dimension analysis, and released code.\n\nRecommendation: engage with it, but tell the authors the evidence is single-shot.","headline":"A sensible method paper whose central superiority claim rests on a single trajectory; worth a referee's time but not yet decisive evidence.","tokens_in":10218,"tokens_out":2051,"would_cite":false,"duration_ms":17583,"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 adding a learned two-dimensional latent state to a probabilistic robot dynamics model captures friction memory well enough to beat physics-based and supervised data-driven friction models in open-loop prediction on…","keywords":["robotics","data-driven modeling","friction identification","probabilistic state-space model","latent variables","expectation-maximization","sequential Monte Carlo","neural network"],"falsifier":"Run the same identification procedure on a second test trajectory whose spectral content and velocity range differ markedly from the training excitation, and refit the model with latent dimension 1, 2, 3, and 4; if the 2D model's open-loop error grows sharply on the new trajectory while higher-dimensional variants hold up, the sufficiency and generality of the two-dimensional Markovian latent state is refuted.","tokens_in":9234,"feed_emoji":"🤖","tokens_out":9956,"duration_ms":72181,"temperature":0.7,"pith_summary":"The paper proposes that the hard-to-model memory effects in robot-joint friction can be captured by augmenting a conventional rigid-body dynamics model with a learned latent dynamic state. Both the friction torque and the evolution of this latent state are neural networks, and the whole probabilistic state-space model is fitted directly from noisy joint-position and motor-current measurements using expectation-maximization with sequential Monte Carlo. On an industrial robot, the resulting model predicts open-loop joint motion with lower error and better stability over a full trajectory than Coulomb-viscous, Stribeck, LuGre, GMS, feedforward neural-network, and recurrent neural-network benchmarks. The practical upshot is that accurate dynamic friction identification no longer requires torque sensors or hand-designed friction laws.","feed_headline":"Hidden-state friction model beats all baselines on a real robot","feed_subtitle":"The learned hidden state keeps open-loop predictions accurate for a full trajectory on an industrial robot.","key_machinery":"The load-bearing object is a probabilistic state-space model whose state is the standard robot state augmented by a latent friction state, $x_t = (q_t,\\dot{q}_t,z_t)$, with the transition given by equation (4). The latent dynamics follow $\\dot{z}_t = \\eta_\\theta(x_t)$ and the friction torque is $\\tau_{f,\\theta}(x_t)$, both parametrized by neural networks, so the extended state carries the memory that static friction laws lack. Identification is performed by maximizing the marginal likelihood via the EM algorithm, with the E-step smoothing distribution approximated by sequential Monte Carlo (particle filtering using 200 particles), yielding the neural-network weights and base inertial parameters directly from noisy encoder data.","core_discovery":"The paper's central claim is that representing unmodeled friction dynamics by a low-dimensional latent state, learned jointly with a neural-network friction model inside a probabilistic state-space model, yields a dynamic robot model that beats both physics-based and supervised data-driven benchmarks in open-loop prediction. Concretely, on the KUKA KR6 R700, the proposed latent-variable model reports a mean squared error of 0.024 and a mean absolute error of 0.11 over the first 10 seconds of open-loop prediction, against best baseline LuGre at 0.027 and 0.12; over the complete test trajectory it reports 0.11 and 0.22, against LuGre at 0.66 and 0.42. The supervised neural network and recurrent network baselines become unstable on longer horizons, while the latent-variable model remains accurate, which the paper attributes to the latent state providing the memory required to stay inside the trained dynamics.","pith_inferences":["A direct check of whether the learned two-dimensional latent state corresponds to a physical quantity (for example, bristle deflection or pre-sliding displacement) would require joint torque sensing, which the current testbed lacks; without such a check, the latent state remains a purely functional quantity.","The latent dimension was set to two without a reported model-selection or identifiability analysis, so a reader should treat it as a tuned hyperparameter rather than a discovered property of the friction physics.","Particle filtering with 200 particles will likely become expensive as the number of joints grows, so scaling this approach to full six- or seven-degree-of-freedom robots probably needs variational or amortized inference rather than plain sequential Monte Carlo."],"forward_implications":["Open-loop dynamic simulation of a robot joint can remain accurate over a full trajectory, not just the first seconds, without joint torque sensors.","The identified latent state provides a ready-built online state estimator that can be reused for friction compensation and output-torque estimation in control.","Because the same maximum-likelihood identification applies to physics-based friction models, the benchmark models could be refitted under the same probabilistic noise model for a more uniform comparison.","Since the latent dynamics carry no physical prior, the same formulation can absorb other unmodeled actuator effects such as backlash or elasticity with no change to the model structure."],"supporting_citations":[{"why":"Supplies the LuGre dynamic friction model, the strongest physics-based benchmark and the inspiration for representing friction memory by a latent state.","marker":"[4]"},{"why":"Supplies the Generalized Maxwell-Slip model benchmark that captures hysteresis and must be outperformed.","marker":"[5]"},{"why":"Provides the recurrent-neural-network dynamic friction baseline whose long-horizon instability motivates the proposed latent-variable model.","marker":"[17]"},{"why":"Represents the static neural-network friction approach that the proposed dynamic model extends and compares against.","marker":"[9]"},{"why":"Establishes the conventional identification protocol and dataset design that the experimental validation follows.","marker":"[19]"},{"why":"Provides the Expectation-Maximization algorithm used to compute the maximum likelihood estimate of the model parameters.","marker":"[22]"},{"why":"Supplies the sequential Monte Carlo filtering and smoothing methods used to approximate the EM expectations.","marker":"[23]"},{"why":"Supplies the Stribeck static friction model used as one of the benchmark baselines.","marker":"[3]"}],"fun_headline_variants":["Latent-state friction model beats all baselines on KUKA","Hidden state friction model wins open-loop on real robot","Latent variable model predicts robot friction better than LuGre","Neural latent state model beats all baselines in friction prediction","Probabilistic latent model learns friction memory for robot joints"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a two-dimensional Markovian latent state, with dynamics learned by a neural network, is identifiable from joint position and motor-current data alone and is sufficient to represent all unmodeled friction dynamics.","fun_headline_variants_meta":{"raw":{"variants":["Latent-state friction model beats all baselines on KUKA","Hidden state friction model wins open-loop on real robot","Latent variable model predicts robot friction better than LuGre","Neural latent state model beats all baselines in friction prediction","Probabilistic latent model learns friction memory for robot joints"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000806,"raw_usage":{"total_tokens":3579,"prompt_tokens":1022,"completion_tokens":2557,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":2474}},"tokens_in":638,"tokens_out":2557,"duration_ms":17150,"temperature":1.0,"reasoning_tokens":2474,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:07:03.258688+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same identification procedure on a second test trajectory whose spectral content and velocity range differ markedly from the training excitation, and refit the model with latent dimension 1, 2, 3, and 4; if the 2D model's open-loop error grows sharply on the new trajectory while higher-dimensional variants hold up, the sufficiency and generality of the two-dimensional Markovian latent state is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LuGre dynamic friction model, the strongest physics-based benchmark and the inspiration for representing friction memory by a latent state."},{"cited_title":"Lampaert, F","cited_arxiv_id":null,"evidence_quote":"Supplies the Generalized Maxwell-Slip model benchmark that captures hysteresis and must be outperformed."},{"cited_title":"Hirose and R","cited_arxiv_id":null,"evidence_quote":"Provides the recurrent-neural-network dynamic friction baseline whose long-horizon instability motivates the proposed latent-variable model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Represents the static neural-network friction approach that the proposed dynamic model extends and compares against."},{"cited_title":"Swevers, W","cited_arxiv_id":null,"evidence_quote":"Establishes the conventional identification protocol and dataset design that the experimental validation follows."},{"cited_title":"McLachlan and T","cited_arxiv_id":null,"evidence_quote":"Provides the Expectation-Maximization algorithm used to compute the maximum likelihood estimate of the model parameters."},{"cited_title":"S :arkk:a and L","cited_arxiv_id":null,"evidence_quote":"Supplies the sequential Monte Carlo filtering and smoothing methods used to approximate the EM expectations."},{"cited_title":"Pennestr `ı, V","cited_arxiv_id":null,"evidence_quote":"Supplies the Stribeck static friction model used as one of the benchmark baselines."}],"review_version":1}