{"id":"922dac08-917a-44d0-8b03-ee5be90afa90","arxiv_id":"2507.16842","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A dual learning framework, RL in simulation plus GAIL from human demonstrations, with a pre-calibrated sim-to-real step, achieves kinematic control of a soft manipulator under loads and in confined pipes.","lead":"The authors built a learning-based controller for a pneumatically driven soft robot arm that keeps working under unknown loads and inside tight spaces, using spring sensors to track the arm's shape. A generalist might read it to see how reinforcement learning and imitation learning combine with simulation-to-real transfer to make soft robots practical in confined environments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The S2R calibration is the load-bearing assumption: it is fit from a small static pose set and cannot transfer the load-dependent saturation boundary, so the disclosed 1 kg failure (Fig. 16) is the expected symptom and the zero-shot robustness claim is bounded by calibration coverage.","rationale":"I read the paper in good faith and found a complete, plausible learning-based control framework with hardware experiments supporting the main architecture. The strongest claim is that SS-ILKC achieves precise path-following and confined-space manipulation under unknown loads with zero-shot deployment. For that claim to hold, the corrected simulator used to train the policy must be an accurate model of the real robot's kinematics and, critically, its actuator saturation boundary. That condition is least secure in the S2R mechanism of Sec. 3.3. The S2R network is trained on a small, unspecified set of static pose correspondences, Eq. (4) is a global linear correction with no stated provenance, and neither component is load-conditioned. The paper's own Fig. 16 demonstrates a 1 kg failure with chamber 2 saturation and 23.76 mm error, which the authors attribute to calibration error near the workspace boundary. That is exactly the failure mode expected when the S2R mapping is inaccurate in under-sampled regions. I do not see a fatal mathematical error or evidence of fraud; the limitation is honestly disclosed. But the disclosed failure is not merely a minor edge case: it directly tests the claimed load robustness and saturation-free operation, and the paper does not quantify the calibration coverage that would make the zero-shot claim reliable. Because the reader's CONDITIONAL verdict already identifies this as the weakest assumption, my stress-test does not move the verdict. I recommend UNCHANGED, with the concrete spatial holdout check above as the cheapest way to settle whether the 1 kg failure is an S2R under-sampling artifact or a more fundamental load-dependence of the saturation boundary.","tokens_in":19213,"tokens_out":3910,"duration_ms":49517,"concrete_test":"Perform a spatial holdout evaluation of NS2R: train the S2R network on all calibration correspondences except those in the high-curvature boundary region where Fig. 16 shows chamber 2 saturating, then compare NS2R predictions against motion-capture ground truth in that held-out region. If the held-out pose error is large and the 1 kg path-following experiment still saturates chamber 2, the S2R under-sampling is confirmed as the cause; if the held-out error is small and saturation persists, the load-dependence of the saturation boundary is the limiting factor and the load-robust claim must be explicitly scoped below 1 kg.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of saturation-free, load-independent, zero-shot deployment rests on the pre-processed sim-to-real mechanism of Sec. 3.3. The RL policy is trained entirely inside a MuJoCo environment whose state is the S2R-corrected pose PS2R = NS2R(P, Lspring), and whose saturation boundary is inherited from the calibrated simulator. Two components of that correction are load-bearing: Eq. (4), Lspring = beta * l + delta, uses two global constants to absorb manufacturing variability, and NS2R is trained by MSE on an unquantified set of static real pose correspondences. Neither component has a load input, and neither is validated with error bars or a spatial coverage analysis. The paper's own 1 kg experiment in Fig. 16 shows precisely the failure mode this assumption predicts: chamber 2 saturates and the tracking error spikes to 23.76 mm near the workspace boundary. The authors attribute this to calibration error near the boundary, which is consistent with the S2R assumption being violated in under-sampled regions. This does not falsify the framework, but it means the claims of 'robust under unknown loads' and 'zero-shot deployment' are only as strong as the calibration coverage, and that coverage is not quantified. The reader's identified weakest assumption is the same one; the additional sharpening is that saturation limits observed at no load may not characterize saturation under a 1 kg load, so the S2R mapping is load-bearing for exactly the saturation-free behavior that the paper highlights.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents SS-ILKC, a learning-based kinematic control framework for a three-module pneumatically actuated soft manipulator. The control operates in sensor space: the policy outputs reference spring-length/frequency commands that a low-level PID controller tracks, with the manipulator state represented by conductive-spring inductance measurements. Training is twofold: a multi-goal reinforcement-learning (TQC) policy is trained in MuJoCo for open-space reaching, and a GAIL-based reward network learns from human demonstrations to handle confined-space tasks such as pick-and-place inside a cross-shaped pipe. A pre-processed sim-to-real (S2R) component, consisting of a linear mapping (Eq. 4) and a lightweight neural network NS2R, corrects simulated end-effector poses to match reality before policy training. Experiments demonstrate circular path following with mean translation error around 11 mm, successful path following with loads up to 500 g using a policy trained without loads, and pick-and-place in a 300 mm pipe that generalizes to a 250 mm pipe. The authors disclose a failure case: a 1 kg load causes chamber 2 to saturate and tracking error to reach 23.76 mm.","tokens_in":19464,"tokens_out":6973,"duration_ms":78496,"significance":"If the claims are substantiated, the framework would be a valuable contribution to soft-robot control: it addresses three hard problems (load-dependent deformation, actuator saturation, and confined-space operation) in a single sensor-space architecture, and the S2R pre-calibration idea is a useful alternative to post-hoc fine-tuning. The paper is honest about its limitations, provides ablation comparisons, and includes a supplemental video. However, the strengths are tempered by the absence of repeated trials and error bars, an S2R calibration whose coverage and error are not quantified, and claims of \"zero-shot,\" \"load-independent,\" and \"saturation-free\" behavior that are stronger than the evidence in Sec. 6 and Sec. 7 supports. The central idea is sound and the experimental scope is appropriate for a robotics venue, but several load-bearing validation points need to be strengthened.","major_comments":[{"comment":"The S2R correction has no load input, and the simulator's saturation boundary is calibrated in unloaded conditions. The 1 kg experiment in Fig. 16 shows chamber 2 saturating near the workspace boundary, with tracking error spiking to 23.76 mm; this is the failure mode one would expect if the S2R mapping does not account for load-dependent changes in the feasible actuation range. The paper attributes the failure to \"calibration errors near the workspace boundary,\" but the more direct interpretation is that the policy commands spring-length references that are infeasible under load. Please either add a load-conditioned term to the S2R mapping or the simulator, or explicitly bound the \"load-independent\" claim to loads below 1 kg and discuss how the saturation boundary scales with load. This is load-bearing because the RL policy is trained entirely inside the S2R-corrected simulation.","section":"§3.3 and §6.2/§7"},{"comment":"No experiment is repeated; all reported path errors, success rates, and load-test results appear to come from single runs. The claims of robustness, generalization, and the relative performance of the ablations (e.g., Fig. 14(c)) cannot be statistically assessed without error bars or confidence intervals. Please report at least three to five repetitions per condition, including the path-following, load, and pipe-transfer experiments, and provide mean and standard deviation for each reported metric.","section":"§6 (all experiments)"},{"comment":"The goal-reaching threshold θ = 0.03, combined with the scaling vector w = diag([0.0056×I3, 0.001×I3]), implies a translation-equivalent tolerance of about 5.36 mm (0.03 / 0.0056). The reported average translation error of 11.29 mm for SS-ILKC is more than twice this tolerance, making it unclear whether the sparse reward Rg = 100 was ever awarded during the path-following evaluation. Please clarify the relationship between the threshold, the scaling vector, and the reported path-following error, and report the fraction of evaluation steps that satisfy the goal condition. This is important for interpreting the learning signal and the meaning of the final error.","section":"§4.2.1 and §6.1"},{"comment":"The GAIL demonstrations are collected from the same task and environment used for evaluation: in Sec. 6.1 the demonstrations are 16 points along the very circular path later used for evaluation, and in Sec. 6.3 the demonstrations are recorded in the same cross-shaped pipe used for testing. This introduces a circularity risk because the reward network may memorize the demonstrated trajectories rather than represent a generalizable task distribution. The 16-to-40 point densification and the 250 mm pipe transfer provide some evidence of generalization, but the paper should either add an evaluation on a path that was never demonstrated or quantify how far the evaluated trajectories lie from the demonstration set. In addition, the number of demonstrations and the number of human demonstrators are not reported; please provide these details.","section":"§5.2 and §6.1/§6.3"},{"comment":"The S2R network NS2R is trained on an unquantified set of real-world pose correspondences. The text states only the optimizer, learning rate, and batch size; it does not report the number of calibration poses, their spatial distribution, the train/validation/test split, or the network's prediction error on held-out poses. Without this information, the zero-shot deployment claim cannot be evaluated, and the risk of the network being accurate only in well-sampled regions remains unaddressed. Please provide a dataset description and a validation error map over the workspace.","section":"§3.3"}],"minor_comments":[{"comment":"The numerator appears to contain a typo: combining Eq. (3) with Lspring = β·l + δ yields a factor of 4π^3, not 4π^4, in front of μ0 N^2 C r^2 f^2. Please verify the derivation and correct the equation.","section":"Eq. (5)"},{"comment":"The values β = 0.93 and δ = 2.45 are given as fixed constants, but the estimation procedure is not described. State how these constants were obtained (e.g., least-squares fit on which dataset) and give their units.","section":"Eq. (4)"},{"comment":"The line \"Sample goal uniformly from state space g ∼ Uniform(st)\" is ambiguous: st is an element of the state space, not a distribution. If the goal is sampled uniformly from a set of possible goals, write this explicitly with the correct notation.","section":"Algorithm 1"},{"comment":"The replay buffer is reset to empty before each discriminator update and then repopulated with new transitions. This is unusual and could destabilize the RL training; please clarify whether this is intentional and discuss the effect on sample efficiency.","section":"Algorithm 1"},{"comment":"The sentence \"the learning framework exhibited a training error of 3.89 mm in the simulation environment\" does not specify which variant of the framework produced this error and under what conditions. Please clarify the context.","section":"§6.1"},{"comment":"The caption mentions that chamber 2 is close to saturation, with a dashed red line, but the saturation pressure level is not defined. Add the saturation threshold value to the figure or caption.","section":"Figure 12(c)"},{"comment":"Della Santina et al. 2020b and 2020c appear to be the same paper; please deduplicate and correct the in-text citations.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is well written and the disclosure of the 1 kg limitation is commendable, but the claims in the abstract and contributions (\"zero-shot,\" \"load-independent,\" \"saturation-free\") go beyond what the presented evidence supports. The main issues—missing error bars, unquantified S2R calibration, and evaluation on the same demonstrated paths—are addressable with additional experiments and careful claim revision. I would not reject the paper; it is a strong candidate after major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a credible systems paper for soft manipulator control. The integration is new—sensor-space multi-goal RL, GAIL rewards learned from physical demonstrations, and a pre-calibrated sim-to-real correction network—and the hardware results are genuinely informative: 40-point circle tracking with ~11 mm error, load tests up to 500 g, and pick-and-place inside a 300 mm cross pipe that transfers to a 250 mm pipe. The internal ablations show S2R and demos both matter. Credit is due for disclosing the 1 kg failure in Fig. 16; that is the paper's most honest moment.\n\nThe soft spots are real but not fatal. 'Zero-shot' is used in a weak sense: real-world data are required for S2R calibration and for the demonstrations that shape the GAIL reward. So 'zero-shot' means 'no online adaptation after pre-calibration.' That is defensible, but the abstract should say so. More importantly, the S2R correction is load-bearing and load-free: Eq. (4) and the network NS2R are fit from static pose correspondences, with no load input, and the policy is trained entirely inside the corrected simulator. The 1 kg experiment is exactly the failure mode this predicts—saturation near the boundary, 23.76 mm error—and the authors' own explanation is calibration error near the workspace boundary. That is consistent, and it does not falsify the framework, but it does bound the robustness claim by calibration coverage, and coverage is not quantified. Also the evaluation is single-run, no error bars, no external baselines, and no code or data released. For an IJRR submission that is a real gap, not a stylistic preference.\n\nThe math and RL machinery are standard and the paper reads as intellectually honest: no evidence of circular reasoning beyond the usual IL concern that demonstrations come from the same tasks used for evaluation, which the load and pipe-size generalization partially mitigate.\n\nBottom line: worth a serious referee. I would accept it for review with the clear expectation that the authors either soften the zero-shot and load-robustness wording or provide a calibration-coverage analysis, and release code/data plus multi-trial results. If those land, the paper is a solid contribution to the soft robotics control literature.","headline":"A credible integration of sensor-space RL, GAIL, and S2R calibration with honest hardware demos; the zero-shot/load-robustness claims run ahead of the evidence because the S2R correction is load-free and statically calibrated.","tokens_in":20087,"tokens_out":2613,"would_cite":true,"duration_ms":30465,"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":"This paper shows that a redundant soft manipulator can be controlled reliably by learning policies in sensor space, so that unknown loads, actuator saturation, and confined-space obstacles can all be handled with zero-shot…","keywords":["soft manipulator control","sensor-space control","reinforcement learning","generative adversarial imitation learning","sim-to-real transfer","actuator saturation","confined-space manipulation","kinematic control"],"falsifier":"Run a dense grid of goal-reaching and circular-path trials near the workspace boundary while loading the gripper with 1 kg and compare the simulated saturation predictions to real chamber pressures. If the real robot saturates in regions where the sim-to-real corrected simulator predicts no saturation, the zero-shot transfer claim is falsified; the paper's own Fig. 16 already shows one such mismatch.","tokens_in":18939,"feed_emoji":"🤖","tokens_out":7004,"duration_ms":77614,"temperature":0.7,"pith_summary":"SS-ILKC is a learning-based kinematic controller for redundant soft manipulators that aims to solve three problems at once: unknown external loads, actuator saturation, and operation in confined spaces. The key idea is to control the robot in sensor space, using conductive springs that measure each chamber's length, so the control objective is load-independent geometry rather than pressure or actuation. An RL-derived multi-goal policy trained in a high-fidelity simulator handles workspace-wide reaching and avoids saturation, while a GAIL-based component learns rewards from sparse human demonstrations for confined-space tasks. A pre-calibrated sim-to-real correction lets the resulting policy deploy zero-shot on real hardware. The authors report path-following errors around 9 to 11 mm and successful pick-and-place inside a cross-shaped pipe, including on a pipe the policy was never trained on.","feed_headline":"Learned sensor-space control guides soft manipulator in confined spaces","feed_subtitle":"Reinforcement plus imitation learning, with a sim-to-real correction, gives zero-shot path following under unknown loads.","key_machinery":"The mechanism that carries the argument is the sensor-space representation: conductive springs integrated along each of the nine chambers act as proprioceptive sensors whose inductance maps to chamber length, so the controller observes geometry instead of pressure. Around this, the framework builds an MDP whose state includes the sim-to-real corrected pose, the goal, and scaled pose error, and whose action is the nine sensor setpoints; a PID loop tracks these setpoints through the pumps. The learning consists of a multi-goal RL stage using Truncated Quantile Critics with hindsight-experience-replay relabeling, and a GAIL stage whose discriminator, trained with a gradient penalty, supplies reward for demonstrations collected by remote control in confined spaces. The pre-processed sim-to-real transfer, a linear correction of spring lengths plus a lightweight network mapping simulated poses to real poses, is what makes the simulator an accurate enough teacher for zero-shot deployment.","core_discovery":"The paper's central claim is that the SS-ILKC framework provides reliable kinematic control for a pneumatically actuated, three-section soft manipulator by replacing actuator-space control with sensor-space control. The state is built from spring-derived sensor feedback and the sim-to-real corrected end-effector pose, the action is the nine chamber sensor signals, and the reward penalizes both reaching error and actuator saturation. Multi-goal RL with hindsight relabeling trains a policy that generalizes across the workspace, and a GAIL discriminator with gradient penalty converts sparse expert demonstrations into reward signals for confined-space behavior. Experiments on the physical robot show average translation errors of about 11 mm on a circular path without loads, about 9 mm under object loads between 35 and 500 g, and successful pick-and-place in a cross-shaped pipe including a generalization run with a narrower 250 mm pipe. The authors state the framework operates without post-training fine-tuning thanks to the pre-processed sim-to-real correction, and they acknowledge saturation near the workspace boundary under a 1 kg load as a residual limitation.","pith_inferences":["Beyond the paper, the same sensor-space plus GAIL plus sim-to-real recipe could be applied to other continuum or soft robots that have embedded proprioceptive length sensing, not just pneumatic three-section arms.","Beyond the paper, a denser, load-aware calibration that samples poses near the workspace boundary under several payloads would likely extend the zero-shot envelope beyond the demonstrated 500 g range.","Beyond the paper, combining the learned reward with online uncertainty estimation about the sim-to-real correction could turn the current open-loop transfer into a safety-aware deployment that knows when it is outside the calibrated region.","Beyond the paper, the GAIL-from-demonstrations component suggests a general way to inject human knowledge into RL controllers for soft robots in environments too complex to simulate faithfully, at the cost of collecting physical demonstrations."],"forward_implications":["If the central claim holds, soft-manipulator kinematic control no longer needs an explicit actuator model or load model: the same policy trained without loads tracks paths while carrying objects up to 500 g.","The pre-calibrated simulator means a trained policy can be deployed zero-shot on real hardware, skipping the usual post-training fine-tuning or online adaptation loop.","Because rewards and demonstrations are expressed in sensor space, the learned behaviors transfer to a more confined environment, with pipe diameter reduced from 300 mm to 250 mm, without retraining the reward network.","Saturation avoidance is encoded during training, so null-space regulation emerges from the learned policy rather than from an optimization step that can push actuators to their limits.","The residual 1 kg-load failure marks the boundary of the method: near the workspace limit, saturation can still defeat the policy."],"supporting_citations":[{"why":"Establishes the sensor-space control concept and the gradient-descent IK solver that the new framework replaces because it saturates on a redundant robot.","marker":"Meng et al. (2024)"},{"why":"Supplies the GAIL objective used to turn expert demonstrations into reward signals for confined-space tasks.","marker":"Ho and Ermon (2016)"},{"why":"Provides the gradient-penalty technique that stabilizes the adversarial reward-network training.","marker":"Gulrajani et al. (2017)"},{"why":"Motivates the hindsight goal relabeling used on both RL transitions and expert demonstrations to improve sample efficiency.","marker":"Andrychowicz et al. (2017)"},{"why":"Frames the multi-goal reinforcement-learning formulation behind workspace-wide policy generalization.","marker":"Plappert et al. (2018)"},{"why":"Supplies the TQC algorithm that stabilizes value estimation during policy optimization.","marker":"Kuznetsov et al. (2020)"},{"why":"Provides the MuJoCo physics engine in which the soft manipulator and its sensors are simulated.","marker":"Todorov et al. (2012)"},{"why":"Represents the model-based alternative for load-dependent control that the learning approach aims to beat.","marker":"Bruder et al. (2021)"},{"why":"Represents prior deep-RL path-following control under varying loads that the framework extends with sensor-space and demonstration learning.","marker":"Centurelli et al. (2022)"}],"fun_headline_variants":["Sensor-space learning steers soft robot in tight spots","Learning from sensors and demos for soft arm control","Zero-shot soft manipulator control via sensor-space learning","Sensor-space RL plus imitation for confined soft robots","Learning robust control for soft manipulators in cramped spaces"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on the assumption that a simulator corrected with a small set of measured poses and a simple spring-length adjustment accurately predicts the real robot's motion and its actuator limits everywhere in the workspace; the 1 kg boundary test shows this can fail where the calibration data are sparse.","fun_headline_variants_meta":{"raw":{"variants":["Sensor-space learning steers soft robot in tight spots","Learning from sensors and demos for soft arm control","Zero-shot soft manipulator control via sensor-space learning","Sensor-space RL plus imitation for confined soft robots","Learning robust control for soft manipulators in cramped spaces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000591,"raw_usage":{"total_tokens":2774,"prompt_tokens":952,"completion_tokens":1822,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":1747}},"tokens_in":568,"tokens_out":1822,"duration_ms":13270,"temperature":1.0,"reasoning_tokens":1747,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:50:07.409918+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a dense grid of goal-reaching and circular-path trials near the workspace boundary while loading the gripper with 1 kg and compare the simulated saturation predictions to real chamber pressures. If the real robot saturates in regions where the sim-to-real corrected simulator predicts no saturation, the zero-shot transfer claim is falsified; the paper's own Fig. 16 already shows one such mismatch.","supporting_citations":[{"cited_title":"IEEE ASME Trans","cited_arxiv_id":null,"evidence_quote":"Establishes the sensor-space control concept and the gradient-descent IK solver that the new framework replaces because it saturates on a redundant robot."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the GAIL objective used to turn expert demonstrations into reward signals for confined-space tasks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the gradient-penalty technique that stabilizes the adversarial reward-network training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the hindsight goal relabeling used on both RL transitions and expert demonstrations to improve sample efficiency."},{"cited_title":"In: CoRL","cited_arxiv_id":null,"evidence_quote":"Supplies the TQC algorithm that stabilizes value estimation during policy optimization."},{"cited_title":"In: IROS","cited_arxiv_id":null,"evidence_quote":"Provides the MuJoCo physics engine in which the soft manipulator and its sensors are simulated."},{"cited_title":"IEEE Robot","cited_arxiv_id":null,"evidence_quote":"Represents the model-based alternative for load-dependent control that the learning approach aims to beat."},{"cited_title":"IEEE Robot","cited_arxiv_id":null,"evidence_quote":"Represents prior deep-RL path-following control under varying loads that the framework extends with sensor-space and demonstration learning."}],"review_version":1}