{"id":"17d2ed48-dbdd-4e67-b55b-54653cc5efca","arxiv_id":"1908.03269","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"RNN and bidirectional-RNN feedforward compensators, trained on measured Baxter motion, reduce trajectory tracking error by roughly 30 to 50 percent versus feedback-only control.","lead":"This paper uses recurrent neural networks learned from actual robot motion to generate feedforward commands that reduce trajectory tracking error on a flexible-joint Baxter robot, improving over feedback-only control by roughly 30 to 50 percent in experiments. The method matters because it removes the need for an explicit dynamic model, and the bidirectional-network version can generate compensation in real time for teleoperation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported improvements may be due to the noncausal preview and feedback, not to a learned inverse: the BRNN is never validated and no trivial lead/lag feedforward baseline is compared.","rationale":"The reader's weakest assumption focuses on whether the BRNN learns a usable inverse, and that is indeed part of the concern. My stress-test sharpens it by pointing out that the experimental protocol conflates the learned inverse with the noncausal preview inherent in the 50-step window. The BRNN has access to future desired setpoints, while the baseline feedback controller does not, so the reported improvement could be a trivial consequence of lead compensation rather than of learning the flexible-joint dynamics. The teleoperation result is especially vulnerable because the future data are obtained by inserting a 0.24 s delay, which changes the comparison baseline. This is a concrete, testable confound: a constant-lead feedforward with the same preview would isolate the contribution of the learned inverse. The ILC derivation gap in Eq. (4) is a secondary but real reproducibility issue. These concerns are addressable by additional experiments and would not, at this stage, justify rejection; they do justify keeping the verdict CONDITIONAL, which is what the reader recommended. I therefore leave the verdict unchanged but broaden the rationale for conditionality.","tokens_in":10431,"tokens_out":13443,"duration_ms":152752,"concrete_test":"On the same test trajectories used in Tables I–IV, replace the BRNN with a constant-lead feedforward qf(t)=qd(t+Δ), with Δ tuned on the training set (e.g., Δ∈{1,...,25} steps), keeping the same feedback gain k and the same 0.24 s delay protocol for teleoperation. If this trivial noncausal baseline achieves at least 80% of the BRNN's l2 error reduction, then the learned inverse is not necessary and the central claim is not supported; if it achieves much less, the neural learning is doing real work.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that learned RNN/BRNN feedforward reduces tracking error by 30–50%. The paper never validates the BRNN inverse separately: no prediction-error numbers on held-out (q, qd) pairs are reported, and no ablation against simpler noncausal feedforwards is given. Because the BRNN is fed a 50-step window of the desired trajectory (±25 steps, Eq. 5), it has a substantial preview unavailable to the baseline feedback controller (Eq. 2), which uses only qd(t+1). In the teleoperation experiment (Sec. V.D), this preview is obtained by delaying the commanded trajectory by 0.24 s, so the comparison in Table IV does not control for latency. A fixed time-advanced feedforward, qf(t)=qd(t+Δ), might achieve comparable improvement without any learned dynamics. If so, the central claim's attribution of the improvement to neural learning of unknown dynamics is not supported. Additionally, Eq. (4)'s ILC update is stated for a linear adjoint G*(s), but no derivation from the trained nonlinear RNN is given, leaving the first approach under-specified and difficult to reproduce.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper addresses trajectory tracking of a flexible-joint Baxter robot whose inner torque servo has unknown dynamics. The authors propose two feedforward compensation schemes: an offline iterative learning control (ILC) that uses a unidirectional RNN trained to predict the forward dynamics, and a direct inversion scheme in which a bidirectional RNN (BRNN) trained on (actual response, commanded setpoint) pairs is used to filter the desired trajectory into a feedforward command. The feedforward is added to a proportional feedback baseline. Experiments on sinusoidal, random joint, Cartesian square, and teleoperated trajectories report a 30-50% reduction in l2 tracking error relative to the baseline. The paper argues that the learned models capture the coupled, flexible-joint dynamics and that the BRNN's bidirectional window captures the noncausality of the inverse system.","tokens_in":10622,"tokens_out":7002,"duration_ms":69868,"significance":"If the improvement is causally attributable to the learned internal dynamics, the paper would provide a practical, data-driven way to improve tracking of collaborative robots that expose only a joint-setpoint interface, without requiring torque-level control or an accurate simulator. The study has genuine strengths: experiments on a real Baxter with series-elastic actuators, joint coupling explicitly demonstrated, a comparison of two learning-based architectures for the same data set, and a large training corpus (500 trajectories, 1.2M samples). It also makes a conceptually interesting use of a BRNN as a noncausal inverse filter. However, the causal attribution is currently under-supported: the BRNN is not validated as an inverse model, the teleoperation comparison does not control for a 0.24 s preview advantage, and the ILC update for the nonlinear RNN is stated only for a linear adjoint. These issues must be resolved before the central claim can be accepted.","major_comments":[{"comment":"The ILC update in Eq. (4) is written for the adjoint G*(s) of a linear time-invariant system, but the forward model used to generate the gradient is the trained nonlinear RNN. The paper does not derive the corresponding adjoint for the RNN, nor does it explain how G*(s)eq is computed (e.g., by backpropagation through time or by linearization). Because the first approach's feedforward command is the output of this ILC, the method is under-specified and not reproducible as it stands. Please provide the gradient computation for the RNN or justify that the linear ILC update applies.","section":"IV.B.3, Eq. (4)"},{"comment":"The BRNN is trained on pairs (q, qd) in which the input is the measured response and the output is the commanded setpoint, but at inference it is applied to desired trajectories qd to produce qf. This presumes the command-to-response map is invertible and that the learned network generalizes from response windows to desired windows. No held-out prediction error for the BRNN is reported, and no ablation is given against a simpler noncausal feedforward (e.g., qf(t)=qd(t+delta)). Since the BRNN sees a +-25-step window (Eq. (5)) whereas the baseline controller (Eq. (2)) uses only qd(t+1), the reported 30-50% improvements may be largely due to preview rather than to a learned inverse. Please validate the inverse model and compare with a matched-preview baseline.","section":"IV.C.1 and V (Tables I-III)"},{"comment":"In the teleoperation experiment, the BRNN requires 24 future user inputs, which the authors obtain by delaying the commanded trajectory by 0.24 s. Table IV compares this delayed, preview-enhanced controller against a baseline feedback controller that does not have access to future setpoints and is not delayed in the same way. The 30% improvement therefore cannot be unambiguously attributed to the learned inverse dynamics. The authors should compare against a baseline that has the same 0.24 s preview (e.g., a time-shifted desired trajectory) or otherwise control for latency.","section":"V.D"}],"minor_comments":[{"comment":"The claim that additional ILC iterations on the physical robot further improve performance is not supported by any reported results; please report these experiments or remove the claim.","section":"V.C"},{"comment":"There is a notational slip: the sentence 'for each input qd and output trajectory q' should read 'for each input q and output trajectory qd', since the BRNN maps response to command. Also, the sample counts 1225000 and 1225500 differ from the formula (2500-T or 2501-T times 500); please check the arithmetic.","section":"IV.C.2"},{"comment":"The choice T=50 is justified only qualitatively; since T determines the preview length and the network memory, a sensitivity study (or at least a statement of the tradeoff) would help the reader understand the role of this hyperparameter.","section":"IV.B.2 and IV.C.2"},{"comment":"The four testing trajectories are hard to distinguish from the training-trajectory background; using distinct colors or markers would improve clarity.","section":"Fig. 4"},{"comment":"All tables report single-run errors without error bars or statistical tests; over repeated trials this would make the improvement claims more robust.","section":"V (Tables I-IV)"}],"recommendation":"major_revision","confidential_remarks":"The main concern is whether the reported improvements are due to learned inverse dynamics or to the inherent noncausal preview. If the authors cannot demonstrate superiority over a simple time-advanced command with matched preview, I would be inclined to lower the rating. The paper is also quite thin for a journal submission; the experimental section would benefit from more trials and a validation of the learned inverse."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You'll want to know about this paper: it reports a head-to-head comparison of two RNN-based feedforward compensation schemes for a Baxter arm with flexible joints, evaluated on a physical robot. The experiments are real and the reported tracking improvements (30–50% reductions in l2 error vs a proportional-only baseline) are substantial. That is the good news.\n\nThe genuinely new bits are the systematic comparison of an offline RNN+ILC scheme with a bidirectional-RNN inverse-dynamics scheme on the same hardware, and the teleoperation demo where the BRNN runs in real time. The authors are also honest about the coupling of the Baxter joints—they train one coupled network across all seven joints rather than per-joint models—and they provide their middleware bridge code. The data-collection effort (500 trajectories of command-response data) is non-trivial.\n\nNow the soft spots. First, the paper reports no repeated trials, no error bars, no statistical significance. For a claimed 30–50% improvement, we need to know how much of that is run-to-run variability. Second, the ILC update in Eq. (4) is written for a linear adjoint G*(s), but the learned forward model is a nonlinear RNN. The authors never say how the adjoint of the RNN is computed or whether the gradient is accurate. That makes the first method under-specified and hard to reproduce. Third, and most important, the BRNN inverse is never validated on its own, and no trivial baseline is tested. Both RNN inputs contain a substantial look-ahead window: a 50-step desired-trajectory window for the forward RNN, and a ±25-step window for the BRNN. The baseline controller only uses qd(t+1). So part of the improvement likely comes from having future reference information, not from learned nonlinear dynamics. In the teleoperation experiment, the authors deliberately insert a 0.24 s delay to gather future inputs, and they do not compare against a simple time-advanced feedforward. That is a real confound. The central claim as written—that neural learning of unknown dynamics drives the improvement—is not fully supported without a preview-control baseline.\n\nThis is a well-written empirical study, and I'd send it to review—a serious referee could push for the missing controls and statistics. It would be a solid contribution if the authors add multi-trial data, a lead/lag feedforward baseline, separate validation of the learned inverse, and a clear description of the ILC gradient computation. As it stands, it is valuable mainly to robotics practitioners who want to see what RNN feedforward looks like on real hardware, but it is not yet a convincing proof that the learned dynamics are the cause.\n\nFor yourself: worth a quick look for the experimental setup, but I would not cite the main quantitative claim until the preview confound is addressed.","headline":"A useful empirical comparison of RNN-based feedforward schemes on a Baxter, but the main claim that learned dynamics drive the improvement is undercut by an untested preview baseline.","tokens_in":11171,"tokens_out":3362,"would_cite":false,"duration_ms":34486,"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":"Learned neural feedforward cuts flexible-joint robot tracking error by 30–50% without a dynamics model.","keywords":["trajectory tracking","flexible-joint robot manipulator","recurrent neural network","bidirectional RNN","iterative learning control","feedforward compensation","inverse dynamics","Baxter robot"],"falsifier":"Record a set of held-out command-response pairs, feed each measured response into the BRNN to get a predicted command, replay that command on the robot, and measure the resulting error; if the replayed output deviates as much as or more than the baseline proportional controller, the learned inverse is not doing the work claimed. A lighter check is to run a test trajectory whose manipulability falls outside the light-blue training region in Fig. 4 and see whether the 30–50% improvement disappears.","tokens_in":10128,"feed_emoji":"🤖","tokens_out":6067,"duration_ms":57709,"temperature":0.7,"pith_summary":"The paper claims that neural networks trained only on observed command–response data can compensate for the unknown, coupled, flexible-joint dynamics of a robot arm. It builds two feedforward controllers for a Baxter arm — one offline iterative-learning scheme using a forward-dynamics RNN, and one direct inverse-dynamics filter using a bidirectional RNN — and reports that both cut tracking error by roughly 30–50% relative to a proportional feedback baseline on unseen trajectories. The point is to show that precise trajectory tracking can be recovered without an analytic robot model or system identification, and that the bidirectional inverse method works in real time for teleoperation.","feed_headline":"Neural feedforward cuts flexible-arm tracking error by 30-50%","feed_subtitle":"Two RNN schemes learn a Baxter arm's hidden joint flexibility from command–response data and beat proportional feedback.","key_machinery":"The carrying object is a pair of recurrent networks sharing one dataset: a 4-layer unidirectional GRU network that maps a 50-step window of commanded joint positions to the joint configuration 50 steps later, and a 2-layer bidirectional GRU network (BRNN) that maps a 50-step window of measured joint positions centered at time $t$ back to the command at time $t$. The BRNN is what makes the inverse-dynamics approach work: because the stable inverse of a strictly proper or non-minimum-phase system is non-causal, each feedforward command must look both backward and forward along the desired trajectory, and the centered window supplies that context. The forward RNN feeds a gradient-based ILC update $u^{k+1} = u^k - \\alpha_k G^*(s) e_q$, where $G^*(s)$ is the adjoint of the learned forward model; the BRNN simply filters the desired trajectory to produce the feedforward command. Both outputs are added to a proportional feedback term to form the commanded joint setpoint.","core_discovery":"On its own terms, the paper's claim is that the inner-loop command-to-response dynamics of a flexible-joint manipulator can be learned directly from recorded command and response trajectories, and the learned map can be used for feedforward compensation in two ways. First, a unidirectional RNN approximates the forward dynamics and supplies the model for an offline gradient-based iterative learning control that refines the command before execution. Second, a bidirectional RNN trained on the same data with output and command swapped approximates the (generally non-causal) inverse dynamics, so the desired trajectory can be filtered directly to produce feedforward commands in real time. Experiments on the Baxter left arm show average reductions in $\\ell^2$ tracking error of more than 50% for sinusoidal and Cartesian trajectories, over 40% for random joint trajectories, and over 30% for teleoperated trajectories, compared with a proportional feedback baseline; the BRNN method is slightly more accurate than RNN-ILC in most of the reported cases.","pith_inferences":["A testable extension is to measure the BRNN's inverse-model prediction error on held-out command–response pairs; if the inverse mapping is not accurate, the feedforward can amplify error instead of reducing it, which the paper does not assess.","The same bidirectional-inversion recipe should transfer to other actuation chains with delay or flexibility—hydraulic actuators, cable-driven arms, or soft robots—where the stable inverse is likewise non-causal.","The reported 0.24 s look-ahead delay at 100 Hz suggests a tunable trade-off: a shorter BRNN window would reduce teleoperation latency but may lose the non-causal information the inverse needs.","Combining the learned feedforward with a more sophisticated feedback law (the paper uses pure proportional gain) is the obvious next step; stability margins of the closed loop with the learned inverse are not analyzed here."],"forward_implications":["A robot's unknown joint-flexibility dynamics can be compensated for from measurement data alone, without torque-level models or system identification.","A single offline-trained BRNN can produce feedforward commands in real time for trajectories that are not known in advance, such as teleoperated motion.","The learned forward RNN makes iterative learning control practical offline, avoiding slow iterations on the physical robot.","The reported reductions in $\\ell^2$ tracking error generalize across unseen sinusoidal, random, and Cartesian trajectories, with further improvement available from additional physical ILC iterations.","Between the two approaches, the inverse-dynamics BRNN is preferred for online use since it needs only a forward pass, while RNN-ILC requires the entire trajectory ahead of time."],"supporting_citations":[{"why":"Defines bidirectional RNNs, the architecture used to capture the non-causal inverse dynamics.","marker":"[9]"},{"why":"Supplies the idea of training a network on command-response pairs with input and output swapped to approximate inverse dynamics.","marker":"[19]"},{"why":"Provides the gradient-based ILC training procedure and the prior NN-plus-ILC method this paper adapts to the flexible-joint Baxter.","marker":"[22]"},{"why":"Supports using recurrent networks to approximate dynamical systems from data.","marker":"[8]"},{"why":"Underlies the MIMO gradient-based ILC update rule used with the forward RNN.","marker":"[26]"},{"why":"Provides the resolved-velocity controller that generates the teleoperated desired trajectories for the real-time BRNN test.","marker":"[27]"}],"fun_headline_variants":["RNN learns flexible-joint dynamics, slashes tracking error 50%","Bidirectional RNN learns non-causal inverse, cuts arm error 40-50%","Neural net learns hidden joint flexibility, boosts tracking 30-50%","Baxter arm learns its own flex via RNN, halves tracking error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme rests on the assumption that the command-to-response map of the robot is invertible over the tested region and that a 50-step bidirectional window plus the training trajectories capture that inverse well enough; the paper never separately verifies the learned inverse's prediction accuracy.","fun_headline_variants_meta":{"raw":{"variants":["RNN learns flexible-joint dynamics, slashes tracking error 50%","Bidirectional RNN learns non-causal inverse, cuts arm error 40-50%","Neural net learns hidden joint flexibility, boosts tracking 30-50%","Baxter arm learns its own flex via RNN, halves tracking error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001321,"raw_usage":{"total_tokens":5419,"prompt_tokens":1023,"completion_tokens":4396,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":639,"completion_tokens_details":{"reasoning_tokens":4310}},"tokens_in":639,"tokens_out":4396,"duration_ms":28809,"temperature":1.0,"reasoning_tokens":4310,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:19:38.421466+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a set of held-out command-response pairs, feed each measured response into the BRNN to get a predicted command, replay that command on the robot, and measure the resulting error; if the replayed output deviates as much as or more than the baseline proportional controller, the learned inverse is not doing the work claimed. A lighter check is to run a test trajectory whose manipulability falls outside the light-blue training region in Fig. 4 and see whether the 30–50% improvement disappears.","supporting_citations":[{"cited_title":"Bidirectional recurrent neural net- works,","cited_arxiv_id":null,"evidence_quote":"Defines bidirectional RNNs, the architecture used to capture the non-causal inverse dynamics."},{"cited_title":"Deep neural networks for improved, impromptu trajectory tracking of quadrotors,","cited_arxiv_id":null,"evidence_quote":"Supplies the idea of training a network on command-response pairs with input and output swapped to approximate inverse dynamics."},{"cited_title":"Industrial Robot Trajectory Tracking Using Multi-Layer Neural Networks Trained by Iterative Learning Control","cited_arxiv_id":"1903.00082","evidence_quote":"Provides the gradient-based ILC training procedure and the prior NN-plus-ILC method this paper adapts to the flexible-joint Baxter."},{"cited_title":"High performance motion tracking control for electronic manufacturing,","cited_arxiv_id":null,"evidence_quote":"Underlies the MIMO gradient-based ILC update rule used with the forward RNN."},{"cited_title":"Software framework for robot-assisted large structure as- sembly,","cited_arxiv_id":null,"evidence_quote":"Provides the resolved-velocity controller that generates the teleoperated desired trajectories for the real-time BRNN test."}],"review_version":1}