{"id":"2e2d2f12-2a97-441d-a914-ff34e35a75f1","arxiv_id":"2504.13582","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A hysteresis-aware neural network with pressure-direction inputs, combined with PPO control trained in a neural-network simulator, tracks trajectories on a soft surgical robot with 0.126 to 0.250 mm error.","lead":"This paper builds a soft surgical robot and trains a neural network model that uses both air pressure and the direction of pressure change to predict the robot's whole-body shape. The authors then use reinforcement learning in a simulator built on that model and report sub-millimeter trajectory tracking on the physical robot.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The quasi-static hysteresis model may not predict soft-robot states at control frequency; the authors' LSTM future-work admission underscores the gap.","rationale":"The paper's strongest claim is that a hysteresis-aware neural network with pressure-direction inputs, used as the transition model in PPO, achieves sub-millimeter real-world tracking. The most load-bearing premise is that the static map Tnn trained from quasi-static data remains accurate at the control loop's timescale. This premise is weakly supported. First, the model is Markovian in current pressure and binary direction, while the physical system has rate-dependent viscoelasticity and memory; the 3-second stabilization delay in data collection ensures the training distribution is equilibrium states, not transient ones. Second, the authors' conclusion explicitly lists temporal modeling (LSTM) as future work, acknowledging the current model lacks dynamics. Third, the real-world evaluation is thin: two trajectories, one run each, no error bars, and the motion capture system is described as 'millimeter-level,' making errors of 0.126–0.250 mm potentially comparable to measurement noise. Fourth, the claimed 84.95% MSE reduction does not match the reported numbers in Section V-A (0.3947 vs. 0.0693 gives 82.4%); while this arithmetic issue is a correctness blemish, the quasi-static transfer issue is more fundamental. A concrete test would be to evaluate the model on dynamic data without the settling delay or to repeat real-world tracking at higher speeds with measurement uncertainty quantified. If the model error grows significantly under dynamic excitation, the RL policy is likely overfitted to the quasi-static simulator, and the reported real-world accuracy cannot be extrapolated to general surgical scenarios. Since the reader's weakest assumption already captures this concern, I agree with the CONDITIONAL verdict, pending dynamic validation and artifact release.","tokens_in":10144,"tokens_out":6644,"duration_ms":59929,"concrete_test":"Collect dynamic data by commanding the real robot through the same circular and square trajectories at the actual control rate, recording pressures and positions without the 3-second settling delay. Evaluate the HAW-NN model's prediction MSE on this dynamic dataset. If the dynamic MSE is substantially higher than the quasi-static test MSE (0.0693), the RL transition function is inaccurate at control frequency. Additionally, report the control step duration and the motion capture system's static repeatability (e.g., standard deviation of a fixed marker) to assess whether the 0.126–0.250 mm real-world errors exceed measurement noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the PPO policy trained in the HAW-NN simulator transfers to real-world tracking relies on the transition function st+1 = Tnn(st, at) being accurate at the control loop's timescale. However, the model is a feedforward MLP trained exclusively on quasi-static samples: each data point follows a 3-second stabilization delay and then 1 second of averaged recording. Its inputs are only current pressures and binary direction signs, with no pressure rate or history. Soft pneumatic robots exhibit rate-dependent viscoelasticity, creep, and transient dynamics, which a static Markovian map cannot represent. The authors implicitly concede this in the conclusion, listing temporal modeling approaches such as LSTM as future work, thereby acknowledging that the current architecture lacks dynamic modeling. If the robot is controlled significantly faster than the quasi-static stabilization time, the simulator's whole-body position predictions will be systematically biased. The real-world evidence—two trajectories, single-run average errors of 0.126–0.250 mm, and a motion capture system described only as 'millimeter-level'—may not expose this bias if the tested trajectories are slow or if measurement noise is comparable to the reported errors. The simulation-to-real error gap (0.0438 to 0.250 mm for the circle; 0.0104 to 0.126 mm for the square) is consistent with a model mismatch. Because the policy is trained entirely within this simulator, any systematic error at control frequency would compromise the validity of the control claims beyond the specific tested trajectories.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper designs a three-chamber, fiber-reinforced pneumatic soft robot for laparoscopic surgery and proposes a hysteresis-aware whole-body neural network (HAW-NN/HWB-NN) that predicts the coordinates of key feature points along the robot body from current chamber pressures plus a binary direction-of-change input. This network is used as the transition model in a parallel PPO training environment, and the resulting policy is deployed on the real robot for circle and square trajectory tracking. The manuscript also reports a two-robot phantom liver ablation demonstration. The headline results are an 84.95% reduction in prediction MSE for the 6D direction-augmented model over a 3D pressure-only model, and real-world tracking errors of 0.126-0.250 mm.","tokens_in":1443,"tokens_out":1515,"duration_ms":62476,"significance":"If the results are confirmed, the paper offers a practically relevant data-driven modeling and whole-body RL control pipeline for soft robots, with a real-world deployment benchmark and a surgical phantom demonstration. The strengths are a head-to-head comparison across six architectures on the same test set, a real-world deployment that provides an external check on the simulator, and a whole-body key-point representation that is useful for tasks beyond endpoint control. The central risks are that the learned transition model is static and quasi-static, the real-world evaluation lacks repeated trials and an appropriate control ablation, and the headline percentage reduction is not consistent with the numbers in the text. These issues are local and correctable, so I view the manuscript as requiring major revision rather than rejection.","major_comments":[{"comment":"The reported 84.95% MSE reduction is not supported by the numbers given in Section V-A. With the 3D pressure-only model reaching a minimum MSE of 0.3947 and the 6D direction-augmented model reaching 0.0693, the reduction is 1 - 0.0693/0.3947 = 0.824, i.e., about 82.4%. Please correct the abstract and text, or state explicitly which metric and formula produce 84.95%.","section":"Abstract and Section V-A"},{"comment":"The transition function in Eq. (9), st+1 = Tnn(st, at), is a static feedforward MLP trained on quasi-static samples: each data point follows a 3-second stabilization delay and then 1 second of averaged recording, and the input is only current pressures plus a binary direction sign. The model contains no pressure-rate, history, velocity, or time dependence, and no control-loop frequency is reported. At the control rate, rate-dependent viscoelasticity and transients are therefore unmodeled; the paper's conclusion explicitly lists LSTM as future work, acknowledging this gap. This is load-bearing for the sim-to-real claim because the simulated tracking errors in Table II are produced by the same network that forms the RL environment, not by an independent dynamic check. Please provide evidence of model accuracy at the control sampling rate, or augment the model with rate/temporal information.","section":"Sections III-B and IV-A"},{"comment":"The real-world tracking results are reported as single average errors per trajectory, with no number of trials, no standard deviation or error bars, and no statistical comparison. The motion capture system is described only as 'millimeter-level', which is comparable to the claimed sub-millimeter errors. Please report trial counts, variances, and the measurement resolution/accuracy, and state whether the reported values are per-point means over the trajectory.","section":"Section V-C and Table II"},{"comment":"There is no control ablation using the pressure-only model. The modeling comparison in Fig. 5 shows that the 6D input achieves lower test MSE, but the deployed policy evaluation is never compared against a policy trained in the pressure-only environment. Without such an ablation, the trajectory-tracking performance cannot be attributed to hysteresis-awareness. Please add a pressure-only control baseline in simulation and, ideally, on the real robot.","section":"Section V-C"}],"minor_comments":[{"comment":"The nomenclature is inconsistent: the model is called HAW-NN in the abstract and contributions but HWB-NN in Sections III, IV, and V. Please unify.","section":"Throughout"},{"comment":"The number of key feature points n is never specified. The text says it is chosen according to resolution requirements and Fig. 6 shows five points, but the value used in modeling and control should be stated.","section":"Section III-A"},{"comment":"The y-axis of Fig. 5 is labeled only 'MSE'. Since the text cites 0.3947 and 0.0693, please provide units (mm^2) and indicate whether this is an endpoint-only or whole-body error.","section":"Fig. 5"},{"comment":"The caption says 'MSEs' but the listed values are given in mm, not mm^2. Please make the units consistent; the quoted numbers appear to be mean Euclidean errors rather than MSEs.","section":"Table II"},{"comment":"The first contribution bullet claims a 'sim-to-real-to-sim process', but no second sim step is described anywhere in Sections IV or V. Only sim-to-real deployment is shown. Please either add the second sim step or remove the claim.","section":"Contributions and Sections IV-V"},{"comment":"The phrase 'average Eulerian error' should be 'average Euclidean error'.","section":"Section V-D"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the headline 84.95% MSE improvement is arithmetically inconsistent with the paper's own numbers (0.3947 vs 0.0693 implies about 82.4%). This is likely a simple correction, but the authors should be asked to fix it and to clarify the metric. The control evaluation also needs repeated trials and a pressure-only baseline before the real-world tracking claims can be properly assessed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. This is a solid engineering integration, not a conceptual breakthrough: the 'hysteresis-aware' trick is feeding the sign of pressure change into an MLP. The real contribution is a complete pipeline—hardware, whole-body data, learned simulator, PPO, sim-to-real deployment, phantom ablation—and the real-world tracking errors are genuinely small. But the headline 84.95% reduction does not match the numbers in the same paragraph (0.3947 vs 0.0693 gives about 82.4%), and the simulator is a quasi-static feedforward map, so the RL policy's validity at control frequency is an open question the paper does not close.\n\nWhat is actually new and good: the whole-body keypoint representation using planar markers and B-spline reconstruction is sensible and enables laser-pointing tasks that end-tip-only models cannot handle. The direction-sign input is cheap and clearly helps on their test set; the architecture sweep is honest within that same dataset. Deploying the same policy on real hardware with sub-millimeter tracking on two trajectories, plus a two-robot endoscopic ablation demo, is real work. Those results deserve credit.\n\nSoft spots, in proportion. The MSE inconsistency is minor but should be fixed. More importantly: no repeated trials, no error bars, no pressure-only control baseline, so we cannot tell how much of the tracking accuracy comes from hysteresis-awareness versus learned dynamics in general. The simulation errors are not independent evidence—the simulator is the fitted model, so those numbers just say the policy chases its own dream. The quasi-static training data (3-second stabilization before each sample) means the model has no rate information; the authors themselves relegate LSTM/temporal modeling to future work. If the real trajectories are slow relative to the robot's settling time, the mismatch will not show. The paper does not report control-loop timing or trajectory speed to rule this out. That is the biggest structural risk.\n\nWho gets value: soft robotics researchers working on data-driven modeling or sim-to-real for continuum robots. It is a useful reference for the direction-sign feature and the whole-body keypoint idea, but I would not take the sim-to-real numbers as proof of generality. A serious referee should ask for the missing baseline, repeated runs, and a clarification of the control-loop timescale. That revision is feasible; the core pipeline has value. So I would send it to review, expecting a substantial revision.","headline":"A useful engineering pipeline with a simple hysteresis feature and real-world demos, undermined by an internally inconsistent headline number and an under-tested quasi-static modeling assumption.","tokens_in":10992,"tokens_out":2518,"would_cite":true,"duration_ms":24283,"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":"Hysteresis-aware neural modeling plus whole-body PPO achieves sub-millimeter tracking on a real soft robot.","keywords":["soft robotics","hysteresis modeling","whole-body modeling","neural network dynamics","reinforcement learning control","PPO","trajectory tracking","surgical laser ablation"],"falsifier":"Run the trained policy on a pressure trajectory that alternates direction rapidly or varies sinusoidally at 0.5-2 Hz while recording motion-capture positions; if the hysteresis-aware network's predictions diverge from measured positions by more than the reported 0.126-0.250 mm, the quasi-static direction-sign model does not cover the closed-loop regime. A simpler check is to hold pressure constant after a fast ramp and look for creep; visible drift over seconds would violate the training assumption.","tokens_in":9963,"feed_emoji":"🤖","tokens_out":7121,"duration_ms":62185,"temperature":0.7,"pith_summary":"The paper tries to establish that a soft surgical robot can be modeled and controlled without analytical material models by making hysteresis a first-class input to a neural network. It adds the direction of pressure change to each chamber's pressure value, trains a multilayer perceptron on whole-body key-point coordinates, and uses that model as the environment for PPO policy training. The claimed payoff is an 84.95% reduction in prediction MSE compared with pressure-only inputs and real-world trajectory tracking errors between 0.126 and 0.250 mm. If these numbers hold, the approach offers a practical data-driven pipeline for whole-body soft-robot control in tasks such as surgical laser ablation.","feed_headline":"Pressure direction cuts soft-robot model error by 85%","feed_subtitle":"Trained in a neural simulator, the policy tracks real circular and square trajectories within 0.25 mm.","key_machinery":"The central object is the Hysteresis-aware Whole-Body Neural Network (HAW-NN): an MLP whose six-dimensional input is per-chamber pressure plus a signed direction flag, and whose output is the 3D coordinates of n key feature points along the body, with the continuous shape reconstructed by B-spline fitting. The direction flag is the mechanism for hysteresis, because it distinguishes pressurization from depressurization paths that reach the same pressure. The model is trained with a motion-range-weighted MSE that prevents the fixed end's tiny motions from being swamped by the free end's large deflections. It then serves as the deterministic transition function $s_{t+1} = T_{nn}(s_t, a_t)$ inside a 64-environment parallel PPO simulator, letting the policy learn whole-body control without a physical robot in the loop.","core_discovery":"The paper's central claim is that the hysteresis of a soft pneumatic robot can be captured in a static neural network simply by including the direction of pressure change (rising, +1, or falling, -1) alongside the current pressure values in each of the three chambers. Trained on 13,824 quasi-static samples of key feature points along the body, this six-input model reaches a test MSE of 0.0693, an 84.95% reduction from the 0.3947 MSE of the pressure-only baseline. The same network is used as the transition function in a 64-environment PPO simulator, and the resulting policy, when deployed on the physical robot, tracks a circular trajectory with 0.250 mm mean error and a square trajectory with 0.126 mm mean error. A whole-body laser-pointing variant of the task is demonstrated on a phantom liver ablation with two coordinated robots.","pith_inferences":["A direct extension would replace the binary direction flag with a short pressure-history window or pressure rate; if performance holds, the flag was already capturing the dominant memory effect, and if it improves, rate dependence matters.","The real-world evaluation covers two trajectories and end-effector error; body-level errors at the internal key points on untrained trajectories would test the whole-body claim more strictly.","The same direction-sign input could be transferred to other actuators with path-dependent behavior, such as tendon-driven or shape-memory-alloy soft robots, at near-zero modeling cost.","The 3-second settle time in data collection suggests the model is a quasi-static map; a testable extension is active data collection during motion to capture viscoelastic transients."],"forward_implications":["A whole-body state representation makes tasks that depend on global body shape, such as aiming a laser mounted in the robot's central channel, trainable in the same framework.","Because the model is trained from measured key-point data, it applies to soft robots with nonuniform materials or attached tools, where PCC and Cosserat assumptions break down.","Training the policy in a parallel neural-network simulator avoids collecting millions of real-world interactions while still transferring to hardware.","The reported real-world tracking errors of 0.250 mm (circle) and 0.126 mm (square) indicate the sim-to-real gap is small for the tested trajectories.","Coordinated multi-robot tasks, such as endoscope viewing combined with laser ablation, can be executed with the same learned-control pipeline."],"supporting_citations":[{"why":"Defines the Piecewise Constant Curvature model that inspires the piecewise key-feature-point body representation.","marker":"[7]"},{"why":"Supplies the evidence that multilayer perceptrons fit nonlinear flexible-robot kinematics, justifying the MLP choice.","marker":"[12]"},{"why":"Represents prior hysteresis modeling for pneumatic actuators that the direction-aware input extends.","marker":"[14]"},{"why":"Demonstrates model-based reinforcement learning for closed-loop control of soft manipulators, the lineage of this simulator approach.","marker":"[18]"},{"why":"Shows reinforcement learning controllers trained in learned environments for soft robots, supporting the learned-simulator design.","marker":"[21]"},{"why":"Formalizes the Markov decision process used to pose whole-body control as a reinforcement learning problem.","marker":"[22]"},{"why":"Provides the PPO algorithm whose clipped surrogate objective trains the control policy.","marker":"[23]"}],"fun_headline_variants":["Hysteresis-aware model cuts soft robot error by 85%","Soft robot learns hysteresis, tracks within 0.25 mm","Whole-body RL control steers soft robot in surgery","Neural network masters soft robot hysteresis for surgery","85% error drop, sub-mm tracking in soft robot surgery"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a static map from pressure values plus a binary direction sign, learned from data taken after a 3-second stabilization delay, predicts the robot's shape well enough while the controller is running, so that effects like creep, pressure lag, and vibration are negligible at tracking speed.","fun_headline_variants_meta":{"raw":{"variants":["Hysteresis-aware model cuts soft robot error by 85%","Soft robot learns hysteresis, tracks within 0.25 mm","Whole-body RL control steers soft robot in surgery","Neural network masters soft robot hysteresis for surgery","85% error drop, sub-mm tracking in soft robot surgery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000674,"raw_usage":{"total_tokens":3074,"prompt_tokens":955,"completion_tokens":2119,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":2037}},"tokens_in":571,"tokens_out":2119,"duration_ms":13821,"temperature":1.0,"reasoning_tokens":2037,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:04:59.243452+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained policy on a pressure trajectory that alternates direction rapidly or varies sinusoidally at 0.5-2 Hz while recording motion-capture positions; if the hysteresis-aware network's predictions diverge from measured positions by more than the reported 0.126-0.250 mm, the quasi-static direction-sign model does not cover the closed-loop regime. A simpler check is to hold pressure constant after a fast ramp and look for creep; visible drift over seconds would violate the training assumption.","supporting_citations":[{"cited_title":"Controlling Soft Robotic Arms Using Hybrid Modelling and Reinforcement Learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the evidence that multilayer perceptrons fit nonlinear flexible-robot kinematics, justifying the MLP choice."},{"cited_title":"A novel hysteresis mod- elling method with improved generalization capability for pneumatic artificial muscles,","cited_arxiv_id":null,"evidence_quote":"Represents prior hysteresis modeling for pneumatic actuators that the direction-aware input extends."},{"cited_title":"Model- Based Reinforcement Learning for Closed-Loop Dynamic Control of Soft Robotic Manipulators,","cited_arxiv_id":null,"evidence_quote":"Demonstrates model-based reinforcement learning for closed-loop control of soft manipulators, the lineage of this simulator approach."},{"cited_title":"Reinforcement Learning Controllers for Soft Robots Using Learned Environments,","cited_arxiv_id":null,"evidence_quote":"Shows reinforcement learning controllers trained in learned environments for soft robots, supporting the learned-simulator design."},{"cited_title":"Markov decision processes,","cited_arxiv_id":null,"evidence_quote":"Formalizes the Markov decision process used to pose whole-body control as a reinforcement learning problem."}],"review_version":1}