{"id":"a8a2ac23-ab0a-4e25-904e-b7b6410e94fe","arxiv_id":"2504.12702","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A spiking neural network with a liquid state machine and delta encoding predicts 7-DOF robotic arm torques, reportedly reducing error by at least 60 percent, and includes an autoregressive input of the previous torque.","lead":"This paper tests a brain-inspired spiking neural network that predicts the torques needed to move a 7-joint robot arm. The authors report large reductions in prediction error and a tracking demo on two 7-DOF platforms, but the comparison may be affected by an extra input the network receives.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 60% error reduction may be an artifact of an input mismatch: the proposed network receives the previous true torque (Fig. 4b, Sec. III-B) while the baselines are not described as receiving it.","rationale":"The reader's weakest assumption is exactly the load-bearing concern: input-feature mismatch between the proposed network and the baselines. My reading of the manuscript confirms and sharpens it. The problem definition (Eq. 5) excludes τ_{t-1}, while the architecture (Fig. 4b) and the control-framework description (Sec. III-B) explicitly feed the previous actual torque into the MLP. Smooth, 100 Hz robot trajectories make τ_{t-1} highly informative for τ_t, so a model that receives it can show large apparent error reductions for reasons unrelated to LSM temporal processing. Without a statement or experiment showing that PC-ESN, GPR, and v-SVR also receive τ_{t-1} (or an ablation without it), the 'at least 60%' average NMSE claim is not yet substantiated. This does not negate the useful iCub dataset release or the credible trajectory-tracking demonstration, but those are separable from the quantitative superiority claim. The correct disposition is the same as the reader's: accept only conditionally, pending a fair-input comparison. Secondary issues (missing error bars, unspecified PSO selection target, 'no manual tuning' overstatement) reinforce the conditional verdict but are not the primary threat to the central claim.","tokens_in":8401,"tokens_out":6757,"duration_ms":67492,"concrete_test":"Run an ablation/fair-comparison on the released iCub dataset (and Baxter if available). Train all four methods in two configurations: (a) baselines receive augmented input [q, ˙q, ¨q, τ_{t-1}], with the proposed model unchanged; (b) the proposed model receives only [q, ˙q, ¨q] (no τ_{t-1}), with baselines unchanged. Recompute per-DOF and average NMSE (Figs. 6-7). Also fit a trivial AR(1) baseline τ_t = τ_{t-1} to quantify the predictability already contained in the previous torque. If the ≥60% margin is not reproduced in both configuration (a) and configuration (b), the headline comparison is confounded by the extra input and must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is the 'at least 60%' lower average NMSE versus PC-ESN, GPR, and v-SVR on Baxter and iCub. The problem statement in Eq. 5 defines inverse dynamics as τ_t = f(x_t)+ε with x_t = [q, ˙q, ¨q], i.e., no previous torque. However, Sec. III-B states that the actual torque applied to the robotic arm is directly encoded and provided to the network to assist in predicting the torque for the subsequent moment, and Fig. 4b shows the MLP combining the previous time step's torque with LSM features. Thus the proposed model has an extra input, τ_{t-1}, which at 100 Hz sampling is strongly predictive of τ_t for smooth robot motion. The paper never states whether PC-ESN, GPR, or v-SVR were given this same autoregressive input. If they were not, the reported 60% improvement is confounded: it may reflect the extra feedback signal rather than the LSM/delta-encoding spiking mechanism. This is the most load-bearing issue because the abstract and conclusions rest on this comparison, and the architecture text itself creates an internal inconsistency with the problem definition.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an embodied neuromorphic control framework for 7-DOF robotic manipulators. It uses a liquid state machine (LSM) with delta spike encoding and an MLP readout to learn the inverse dynamics mapping from joint state (positions, velocities, accelerations) to joint torques. The authors report experiments on the Baxter robot and the iCub humanoid simulator, claiming torque prediction error reduction of at least 60% relative to PC-ESN, GPR, and v-SVR, together with successful trajectory tracking on the iCub. The paper also introduces a publicly available iCub inverse dynamics dataset and uses particle swarm optimization to search over LSM structural hyperparameters.","tokens_in":8646,"tokens_out":3567,"duration_ms":37299,"significance":"If the empirical claims hold, the paper would be a useful step demonstrating that spiking LSMs can scale to a 7-DOF inverse dynamics problem and control task, and the released dataset is a positive contribution. The combination of delta coding, an LSM, and an MLP readout is a plausible and reasonably novel architecture for exploiting temporal continuity. Strengths include the public dataset, the use of two robotic platforms, and the comparison with multiple baselines and with alternative spike encodings. However, the central quantitative claim is currently undermined by a likely input mismatch between the proposed model and the baselines, and by the lack of aggregate statistics and statistical tests, as detailed below.","major_comments":[{"comment":"The proposed network receives the previous actual torque τ(t−1) as an input to the MLP (Section III-B: “the actual torque applied to the robotic arm is directly encoded and provided to the network to assist in predicting the torque for the subsequent moment”; Fig. 4b caption: “The MLP combines the torque value applied to the robotic arm in the previous time step to predict the torque”), whereas the problem is defined in Eq. (5) as τ = f(x) + ε with x = [q, ˙q, ¨q] and no autoregressive torque term. The manuscript never states whether PC-ESN, GPR, and v-SVR were given this same τ(t−1) input. Since robot torques at 100 Hz sampling are highly autocorrelated, the reported “at least 60%” error reduction may be due to the extra input feature rather than to the LSM/delta-encoding spiking mechanism. The authors must either (a) provide the same τ(t−1) input to all baselines, (b) run an ablation of the proposed model without τ(t−1), or (c) clearly justify and re-frame the comparison accordingly. This is the load-bearing issue for the abstract's central claim.","section":"III-B, Fig. 4b, Eq. (5)"},{"comment":"The abstract and conclusions claim a reduction in average NMSE of at least 60%, but Figs. 6 and 7 show only per-DOF NMSE values without error bars, number of trials, or an aggregate statistic. Table II reports MSE for different encoding schemes, but this is not reconciled with the NMSE figures, and the averaging procedure behind the “average NMSE” claim is not described. To support the central quantitative claim, the paper should report mean and standard deviation over multiple independent runs, specify exactly how the per-DOF, per-platform errors are combined, and include a statistical significance test or at least confidence intervals.","section":"IV-B, Figs. 6–7, Table II"},{"comment":"The PSO procedure searches the LSM structural hyperparameters n, P_input, C_ee, C_ei, C_ie, and C_ii included in Table I, but the objective function used for the search and the data split on which that objective is evaluated are not specified. If the test set or the final validation set influenced the PSO search, the reported generalization results are circular. The authors should state explicitly that PSO was applied only to a training/validation split, describe the objective, and provide this detail in the experimental section.","section":"III-C, Table I, Fig. 9"},{"comment":"The contribution states that the method “eliminate[s] manual parameters tuning,” but the framework still requires the user to set the spike scaling hyperparameter θ in Eq. (7), the PSO search ranges in Table I, the number of liquid neurons, and the MLP training hyperparameters. This claim is overstated and should be reframed as reducing manual tuning or replacing direct search with automated PSO-based selection.","section":"I, III-C, Eq. (7)"}],"minor_comments":[{"comment":"The domain in Eq. (5), R3N×17, appears to be a typo; the intended domain is likely R^{3N} or R^{3N×1}. Please correct the notation.","section":"III-A, Eq. (5)"},{"comment":"The time constants τ1, τ2, and τ3 are not defined, and the use of τ both for these constants and for the torque vector in Eq. (4) is confusing. Please use distinct symbols, such as τ_m, τ_s, and τ_i.","section":"II-A, Eqs. (1)–(2)"},{"comment":"Table II lists a dataset labeled “BaxterRand,” but the text never defines this dataset or explains how it differs from the Baxter experiments described in Section IV-B. The origin and constitution of this dataset should be described.","section":"IV-A, Table II"},{"comment":"The experiments on iCub appear to be performed in the Gazebo simulator rather than on the physical robot. This limitation should be stated explicitly in the abstract and conclusions, where the phrase “two robotic platforms” might be misread as two physical robots.","section":"IV-A, Conclusions"},{"comment":"The y-axis of Fig. 9 is labeled “Average NMSE,” but it is unclear whether this is training error, validation error, or test error, and how the average is computed across joints. Please clarify the reporting protocol.","section":"IV-D, Fig. 9"},{"comment":"The Conclusions contain a typo: “targel-tracking” should be “target-tracking.” Also, the phrase “reduce the average NMSE by 36%” in the Conclusions refers to the encoding comparison, but Table II reports MSE, not NMSE; please use consistent metrics throughout.","section":"V, Conclusions"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising idea and useful empirical work, but the central quantitative claim currently rests on an incompletely specified comparison and lacks aggregate statistics. The most serious concern is the unacknowledged autoregressive torque input in the proposed architecture. If the authors can resolve this by either matching inputs across methods or providing an ablation, and if they add proper error bars and statistical reporting, the paper could be suitable for publication. If the input mismatch cannot be resolved, the 60% claim should be withdrawn or substantially weakened. Please also ensure the public dataset link is stable and that the iCub simulation-only nature is clearly disclosed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know two things about arXiv:2504.12702. First, it is a genuine engineering step: it moves spiking-network inverse dynamics from 2-3 DOF to 7-DOF, releases a public iCub proprioceptive dataset, and shows a plausible torque-tracking demo. Second, the headline '60% error reduction' is probably inflated by an input mismatch. The proposed architecture (Fig. 4b, Sec. III-B) feeds the previous applied torque into the MLP along with LSM features, but the paper never states whether PC-ESN, GPR, or v-SVR get the same autoregressive signal. At 100 Hz, the previous torque is a strong predictor, so the improvement may come from that extra input rather than from spike-based encoding. That is not a peripheral issue; the abstract and conclusions rest on this comparison.\n\nWhat is actually new: the combination of delta encoding, an LSM reservoir, and an MLP readout that conditions on tau_{t-1} is specific to this paper. Scaling SNN inverse dynamics to 7-DOF is non-obvious, and getting it to work on two platforms is worth something. The dataset release is a practical contribution.\n\nThe soft spots are real but mostly fixable. The baseline input mismatch is the load-bearing problem; a fair comparison with identical features would settle it. The paper also lacks aggregate statistics: Figs. 6 and 7 show per-DOF NMSE with no error bars, and Table II's MSE numbers do not reconcile with the NMSE claims. The PSO search is described as being for 'trajectory tracking tasks' but it is not clear whether the search used training or test data; tuning on the test set would be a serious flaw. The claim of eliminating manual parameter tuning is overstated: PSO is still automated tuning. And the delta-coding citation [25] points to a genetic algorithm paper, not spike encoding; that is sloppy.\n\nThe control framework itself and the tracking demo look credible. The weakness is in the evidence for the quantitative claim. This paper deserves a serious referee who can sort out baseline fairness. I would not cite it in its current form, but I would track a revised version.","headline":"A credible 7-DOF neuromorphic control result whose headline 60% error reduction is likely confounded by the proposed network receiving the previous true torque while baselines do not.","tokens_in":9226,"tokens_out":4736,"would_cite":false,"duration_ms":45116,"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":"An LSM-based spiking neural network with delta encoding solves the 7-DOF inverse dynamics problem, cutting average torque-prediction error by at least 60%.","keywords":["spiking neural network","liquid state machine","inverse dynamics","neuromorphic control","robot manipulator","delta encoding","torque prediction","embodied intelligence"],"falsifier":"A re-run of the Baxter and iCub experiments where PC-ESN, GPR, and v-SVR are also given the previous torque as an input feature; if the 60% error reduction shrinks substantially or disappears, the central quantitative claim fails. Comparing the LSM-SNN with an MLP that receives the same delta-encoded features but no spiking dynamics would similarly isolate the contribution of the liquid state machine.","tokens_in":8182,"feed_emoji":"🤖","tokens_out":4699,"duration_ms":38398,"temperature":0.7,"pith_summary":"The paper tries to establish that a spiking neural network built around a liquid state machine, fed with delta-encoded joint motion, can learn the inverse dynamics of a 7-degree-of-freedom robot arm well enough to predict required torques more accurately than three standard machine-learning baselines and to track a target trajectory on two simulated robot platforms. The authors claim this advances embodied neuromorphic control beyond the two- and three-degree-of-freedom demonstrations in earlier work. They also claim delta encoding itself cuts torque-prediction error by roughly a third, and they release a public iCub inverse-dynamics dataset. If true, the result matters because it suggests neuromorphic controllers can handle realistic high-dimensional manipulation tasks without manual parameter tuning.","feed_headline":"Spiking network cuts 7-DOF robot torque error by 60%","feed_subtitle":"Liquid-state spiking control beats PC-ESN, GPR, and v-SVR on Baxter and iCub arms.","key_machinery":"The central object is the Liquid State Machine (LSM), a recurrent network of spiking neurons with random, fixed connections whose internal states act as a spatiotemporal feature extractor. Input joint positions, velocities, and accelerations are first converted to spike trains by delta encoding, with each input dimension mapped to two neurons that fire on positive or negative change, then fed into the liquid layer; the readout is an MLP that also receives the previous actual torque as input. The LSM retains historical state through recurrent connections, and its structure is tuned by particle swarm optimization at initialization. This machinery is what captures the temporal continuity of motion data that the paper argues is missing in feed-forward and sample-independent baselines.","core_discovery":"The paper's central claim is that an LSM-based SNN with delta encoding solves the 7-DOF inverse dynamics problem and produces more accurate torque predictions than PC-ESN, GPR, and v-SVR, with an average NMSE reduction of at least 60% on two platforms, Baxter and iCub. The authors also claim the delta encoding scheme, which fires spikes only on change in the input signal, outperforms rate and direct encoding by up to 36% in mean squared error. A closed-loop trajectory tracking experiment on the iCub simulator shows the predicted torques drive the arm along the desired trajectory.","pith_inferences":["The reported 60% reduction may be partly attributable to the fact that the proposed architecture feeds the previous actual torque into the MLP readout; the paper does not state whether PC-ESN, GPR, or v-SVR also receive this autoregressive input.","A fair comparison would give all methods the same input features, which would isolate whether the gain comes from the spiking network or from the extra torque signal.","If delta encoding plus LSM proves robust across more platforms and physical robots, a natural extension is to run the same network on neuromorphic hardware, where the event-driven property could yield real energy savings.","The released iCub dataset could become a standard benchmark for inverse-dynamics learning on high-DOF arms, enabling other spiking and non-spiking methods to be compared on identical data."],"forward_implications":["Torque prediction error on 7-DOF arms can be reduced by at least 60% relative to PC-ESN, GPR, and v-SVR when using an LSM-based spiking network with delta encoding.","Delta encoding of the joint state outperforms rate and direct encoding on inverse-dynamics prediction, with up to 36% lower MSE.","The LSM structure and connection probability can be automatically searched by PSO, removing the need for manual tuning of network parameters.","The same framework can be deployed on two different 7-DOF platforms, Baxter and iCub, and used for closed-loop trajectory tracking."],"supporting_citations":[{"why":"Supplies the liquid state machine architecture used as the core of the network.","marker":"[10]"},{"why":"Supplies the delta encoding scheme that converts joint-state changes into spike trains.","marker":"[25]"},{"why":"Provides the PC-ESN baseline that the proposed method is compared against.","marker":"[19]"},{"why":"Supplies the GPR and SVR baselines and the established inverse-dynamics learning problem setup.","marker":"[15]"},{"why":"Supplies the iCub humanoid robot platform and the dataset source used for validation.","marker":"[12]"},{"why":"Supplies the Baxter robot as the second test platform.","marker":"[3]"},{"why":"Provides the LSTM-based inverse dynamics learning work that motivates the temporal-continuity argument.","marker":"[22]"},{"why":"Provides the prior closed-loop spiking control work on iCub that this paper extends.","marker":"[26]"}],"fun_headline_variants":["Delta-encoded SNN cuts 7-DOF torque error by 60%","Spiking network drives 7-DOF arms with 60% less error","Neuromorphic control: 60% torque error reduction on 7-DOF","LSM spiking control beats classical models on 7-DOF arms","7-DOF inverse dynamics solved by spiking network with 60% gain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison with baselines assumes that all methods receive the same input features; in particular, the paper does not state whether PC-ESN, GPR, and v-SVR also receive the previous actual torque that the proposed architecture feeds to its MLP readout, so the 60% error reduction may be due to the extra input signal rather than to the spiking network.","fun_headline_variants_meta":{"raw":{"variants":["Delta-encoded SNN cuts 7-DOF torque error by 60%","Spiking network drives 7-DOF arms with 60% less error","Neuromorphic control: 60% torque error reduction on 7-DOF","LSM spiking control beats classical models on 7-DOF arms","7-DOF inverse dynamics solved by spiking network with 60% gain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"total_tokens":1245,"prompt_tokens":896,"completion_tokens":349,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":247}},"tokens_in":512,"tokens_out":349,"duration_ms":3627,"temperature":1.0,"reasoning_tokens":247,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:24:53.278129+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A re-run of the Baxter and iCub experiments where PC-ESN, GPR, and v-SVR are also given the previous torque as an input feature; if the 60% error reduction shrinks substantially or disappears, the central quantitative claim fails. Comparing the LSM-SNN with an MLP that receives the same delta-encoded features but no spiking dynamics would similarly isolate the contribution of the liquid state machine.","supporting_citations":[{"cited_title":"Neural computation 14(11), 2531–2560 (2002)","cited_arxiv_id":null,"evidence_quote":"Supplies the liquid state machine architecture used as the core of the network."},{"cited_title":"In: ICGA","cited_arxiv_id":null,"evidence_quote":"Supplies the delta encoding scheme that converts joint-state changes into spike trains."},{"cited_title":"In: 2015 IEEE/RSJ Inter- national Conference on Intelligent Robots and Systems (IROS)","cited_arxiv_id":null,"evidence_quote":"Provides the PC-ESN baseline that the proposed method is compared against."},{"cited_title":"In: European symposium on artificial neural networks (2008)","cited_arxiv_id":null,"evidence_quote":"Supplies the GPR and SVR baselines and the established inverse-dynamics learning problem setup."},{"cited_title":"In: Proceedings of the 8th workshop on performance metrics for intelligent systems","cited_arxiv_id":null,"evidence_quote":"Supplies the iCub humanoid robot platform and the dataset source used for validation."},{"cited_title":"In: 2013 IEEE conference on technologies for practical robot applications (TePRA)","cited_arxiv_id":null,"evidence_quote":"Supplies the Baxter robot as the second test platform."},{"cited_title":"In: 2017 IEEE- RAS 17th International Conference on Humanoid Robotics (Hu- manoids)","cited_arxiv_id":null,"evidence_quote":"Provides the LSTM-based inverse dynamics learning work that motivates the temporal-continuity argument."},{"cited_title":"IEEE Journal on Emerging and Selected Topics in Circuits and Systems 10(4), 546–556 (2020)","cited_arxiv_id":null,"evidence_quote":"Provides the prior closed-loop spiking control work on iCub that this paper extends."}],"review_version":1}