{"id":"4e97ba91-5c17-471c-8d20-fffcc81189fd","arxiv_id":"2507.21496","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A single linear readout trained on two target signals makes a simulated tensegrity robot multistable, producing two trained and five untrained behavioral attractors that depend on which face touches the ground.","lead":"A simulated soft tensegrity robot was taught to switch between multiple behaviors, such as crawling and staying put, using one simple learned controller and no behavior-specific modules. The same controller also produced several unplanned behaviors, revealing how the robot's shape and ground contact shape its dynamical repertoire.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on a single GA-selected target pair and one stochastic run; without replication it is unclear whether the reported multifunctionality is a robust property or a favorable draw.","rationale":"The reader's verdict is CONDITIONAL, with the weakest assumption identified as MuJoCo fidelity. That is a legitimate limitation, and the paper's own Appendix A acknowledges idealized tendon behavior and instantaneous actuator restlength changes, so it is not an unfair concern. However, I do not regard simulator fidelity as the most load-bearing issue for the central claim: the paper explicitly frames itself as a simulation-based feasibility study, so a hardware-independent proof of concept can still be valuable even before physical validation. The more damaging gap is internal: every headline result in Sections IV B-IV F flows from one phenotype selected by a stochastic genetic algorithm whose third objective explicitly optimizes for closed-loop learnability. The conclusion then states a general condition ('as long as their dynamical trajectories are separated') that the experiments never systematically test. Repeating the pipeline with multiple seeds and with non-evolved target pairs would directly test whether the observed multistability is a robust property of the method or an artifact of a particular optimized target pair. This concern does not overturn the paper's existence proof, but it does support the CONDITIONAL verdict: the central claim should be accepted as a simulation proof of concept while the general design principle remains unverified. Since the reader's verdict already captures this status, the verdict should remain UNCHANGED.","tokens_in":31500,"tokens_out":6441,"duration_ms":87336,"concrete_test":"Rerun the full pipeline from Section III B-C with at least five independent NSGA-II seeds, using the same population size and parameter ranges, and for each resulting phenotype repeat the 20-face initialization sweep of Section IV C plus the closed-loop A/B reconstructions. Then report, across seeds, the success rate for reconstructing both target attractors and whether an attractor repertoire comparable to Attractors A-G with the bottom-face mapping recurs. If the seven-attractor structure and the bottom-face-to-attractor correspondence do not recur in the majority of seeds, the reported multifunctionality is a single optimized realization rather than a general consequence of trajectory separation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Conclusion states that 'the same trained weights of the readout layer can be re-used to control multiple robot behaviors which are coexisting attractors, as long as their dynamical trajectories are separated in the high-dimensional state space.' The evidence for this claim, however, is one phenotype from Generation 73 of a single NSGA-II run (Table I and Section IV B). The genetic algorithm's third fitness objective F3 explicitly optimizes the target signals for learnability in closed-loop (Section III C), so the chosen pair is not an independent sample from the space of possible targets; it is selected for exactly the property the conclusion treats as a sufficient condition. All subsequent claims about untrained attractors (Section IV C), basin structure (Section IV D), post-learning bifurcations (Section IV E), and regions of multifunctionality (Section IV F) inherit this specificity, since they all use the single trained readout W from that one phenotype. This is load-bearing because the paper's headline conceptual contribution is that separation of dynamical trajectories alone permits reuse of one readout for multiple behaviors. As presented, the empirical support for that generalization is one favorable draw from a stochastic search, with no independent GA seeds, no repeated training, and no test on non-evolved target pairs. If the result is a lucky or fragile combination of morphology, target signals, and regularization, then the general design principle is not established, even though the particular example remains an existence proof. The absence of error bars or replication across Section IV therefore weakens the strongest claim more than the acknowledged simulation idealizations do; the latter are explicit limitations, while the former silently determines how far the results can be trusted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes multifunctional physical reservoir computing (MF-PRC) on a simulated six-bar tensegrity robot. Two Lissajous motor-signal targets are optimized with NSGA-II; a single linear readout is trained by ridge regression on 48-dimensional passive-tendon measurements from open-loop simulations; in closed loop, the same readout is claimed to reproduce both target behaviors from their respective initial states. The authors then characterize the resulting multistable system: attractors from 20 bottom-face initial conditions (two trained and five untrained), basin structure in a (p,q)-interpolated input plane, post-learning and pre-learning bifurcations under tendon-stiffness changes, and regions of multifunctionality. The central claim, stated in Section VI, is that one readout can control multiple coexisting behaviors whenever their dynamical trajectories are separated in state space.","tokens_in":31885,"tokens_out":6480,"duration_ms":75725,"significance":"The paper gives a transparent simulation proof-of-concept: physical parameters (Table II), evolutionary settings (Appendix B), classification thresholds (Appendix C), and sensor-level basin data (Appendix D) are all reported, which is a real strength. If the central claim is confirmed, the work would be a useful step toward exploiting body dynamics for multifunctional robot control and for studying untrained attractors in embodied systems. Its significance is currently limited by the single selected phenotype and by the heuristic nature of the attractor classification; the general principle in Section VI goes beyond what is demonstrated.","major_comments":[{"comment":"The strongest claim in Section VI rests on a single phenotype from Generation 73 of one NSGA-II run (Table I). Because the third fitness objective F3 explicitly rewards closed-loop reconstruction accuracy of the chosen targets, the selected pair is not an independent sample from the space of possible targets; it is a favorable draw from a search whose objectives include learnability. All subsequent analyses in Sections IV C through IV F use the same trained readout W from that phenotype. To support the general statement that \"the same trained weights of the readout layer can be re-used ... as long as their dynamical trajectories are separated in the high-dimensional state space,\" the authors should either provide multiple independent GA seeds, repeated training runs, or tests on non-evolved target pairs, or explicitly restrict the conclusion to a case study.","section":"III C, IV B, VI"},{"comment":"Attractor G is classified as non-periodic and \"likely to be a chaotic attractor\" solely from a heuristic autocorrelation threshold (δ_ACF = 0.95) and from the absence of distinct peaks in the power spectrum. This is insufficient to establish chaos or even non-periodicity on a finite-time series; a long-period quasi-periodic orbit or a long transient could satisfy the same criteria. Please add quantitative diagnostics such as a maximal Lyapunov exponent estimate or the 0-1 test for chaos, or relabel the category as \"unclassified non-periodic.\" This matters because Section IV F later refers to \"strange attractors\" in the parameter space.","section":"Appendix C, IV C"},{"comment":"The bifurcation statements in Section IV E and the \"regions of multifunctionality\" in Section IV F are obtained by classifying deterministic single runs at discrete parameter values using the hand-set thresholds of Appendix C. Since the system is deterministic, repeated runs would not add sampling noise, but the classification boundaries themselves are threshold-dependent, and no sensitivity analysis over δ_E and δ_ACF is reported. For example, the identification of a supercritical Hopf bifurcation at k_act ≈ 372.5 N/m in Fig. 8(b) relies on distinguishing a stable fixed point from a small periodic orbit, which is exactly where the threshold choice matters. Please either show that the reported bifurcation locations are robust to reasonable threshold variations or present them as qualitative observations rather than quantitative bifurcation points.","section":"IV E, IV F"}],"minor_comments":[{"comment":"The manuscript says a phenotype from Generation 73 is picked as a representative case, but it does not state the selection rule among the many phenotypes in the final generations; specifying the criterion (e.g., rank on F3, Pareto-front position) would make the case study more reproducible.","section":"IV A, IV B"},{"comment":"There are copyediting errors in this appendix: \"cateogized,\" \"categtorization,\" and \"aformentioned\" should be corrected, and the acronym ACF is defined twice.","section":"Appendix C"},{"comment":"The statement that \"the two-dimensional output dynamics is a good representation of the entire system dynamics because the current system is globally coupled\" is an assertion; a supporting argument or citation is needed, otherwise this should be phrased as a working assumption.","section":"IV C"},{"comment":"Please clarify that the equivalence between the basin boundary and the boundary of bottom faces is established only for the sampled (p,q) region and not for the full high-dimensional state space, since the current wording could be read as a global statement.","section":"IV D"},{"comment":"The data availability statement is vague; given the complexity of the MuJoCo simulation and the evolutionary search, releasing the simulation model and the optimization code would substantially aid reproducibility.","section":"Data Availability"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a robotics or physical-computing venue, and the single-case demonstration is interesting. The requested revisions are feasible incremental experiments rather than a change of concept: repeating the genetic algorithm with multiple seeds, adding quantitative chaos diagnostics, and checking threshold sensitivity. I do not see a basis for rejection, but the general principle stated in the conclusion should not be accepted on the current single-draw evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is the first real attempt I know of to put multifunctional reservoir computing on a physical reservoir in a soft tensegrity robot, and the core demonstration holds up as an existence proof. A single linear readout, trained on open-loop data from two target signals, makes the closed-loop robot multistable: it crawls from one set of initial conditions, oscillates from another, and shows five untrained attractors that line up with which icosahedron face is on the ground. That bottom-face organization is genuinely interesting and gives the morphology a clear role in shaping the attractor landscape. The paper is also well organized, the simulation parameters are reported in enough detail to replicate, and the authors are honest about idealizations like tendons that push as well as pull and instantaneous restlength changes.\n\nThe soft spots are real but not fatal for a proof of concept, and they cluster around one issue: almost everything after Section IV B depends on one phenotype from one NSGA-II run. The genetic algorithm's third objective explicitly selects target pairs for learnability, so the chosen pair is not an independent draw from the space of possible targets. That means the conclusion's claim—that one readout can be reused as long as trajectories are separated in state space—is presently supported by a single favorable case, not a scan over targets or seeds. There are also no error bars, no repeated runs, and no code or data shipped. The attractor classification is threshold-based, which is fine for a survey of the landscape, and I note the authors label attractor G as \"likely chaotic\" rather than proven, which is appropriate. None of this invalidates the central result; it just limits how far the general principle can be trusted.\n\nWho gets value from this paper: people working in physical reservoir computing, soft robotics control, and embodied intelligence. It is a solid existence proof with a useful discussion of untrained attractors and pre/post-learning bifurcations in a physical system. I would not desk-reject it, but I would send it to review with a clear request: show at least a few independent GA seeds or repeated trainings, report variance, and ideally release the simulation code and data. The single-case issue should be the main thread for the reviewers.\n\nMy take: deserves serious peer review with heavy revision. The idea is timely and the execution is clean; the support just needs to be a little broader before the general claims are established.","headline":"A credible simulation proof-of-concept that one readout can hold multiple behaviors in a soft tensegrity robot, but the general claim rests on a single GA-selected case and needs replication.","tokens_in":32348,"tokens_out":1549,"would_cite":true,"duration_ms":21555,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper shows that one trained linear readout can turn a simulated soft tensegrity robot into a multistable system whose behaviors are selected by which face of the icosahedral body touches the ground.","keywords":["tensegrity robot","physical reservoir computing","multifunctional reservoir computing","multistability","attractor reconstruction","untrained attractors","soft robotics","locomotion control"],"falsifier":"Build a physical six-bar tensegrity robot with the same geometry, tendon stiffness, damping, and actuator placement, run the identical open-loop training with the two Lissajous targets, and close the loop from the two final states. The central claim fails if the crawl and the on-the-spot oscillation do not reappear, or if the reported face-to-attractor map (for example, face 17 giving crawling and face 13 giving oscillation) is not reproducible across trials.","tokens_in":31274,"feed_emoji":"🤖","tokens_out":7805,"duration_ms":82288,"temperature":0.7,"pith_summary":"The paper attempts to establish that multifunctional reservoir computing, previously shown in artificial neural networks, can be carried by the physical body of a soft tensegrity robot. The authors train one linear readout on tendon-length-and-velocity measurements collected while the simulated robot is driven by two different Lissajous motor signals, then close the feedback loop. They find that the same readout reproduces both target behaviors from the two open-loop final states, so the robot-environment system becomes multistable with two trained attractors. Starting the closed-loop system from twenty resting initial poses reveals five additional untrained attractors, including a fixed point, periodic orbits, and a chaotic-looking tremor, whose identity is selected mainly by which face of the icosahedral robot touches the ground. If the claim holds, a soft robot can exhibit multiple behaviors with one controller and switch between them by changing its body configuration or applying a perturbation.","feed_headline":"One readout turns a soft robot into seven behaviors","feed_subtitle":"Simulated tensegrity robot switches among crawling, swaying, and still states by changing which face touches the ground.","key_machinery":"The central object is the closed-loop feedback law $u(t+\\tau)=W r(t)$, where $r(t)\\in\\mathbb{R}^{48}$ holds the lengths and velocities of the 24 passive tendons and $W\\in\\mathbb{R}^{2\\times 48}$ is obtained by ridge regression over concatenated open-loop data from two simulations. The identity doing the work is that two behaviors' trajectories are separated in this high-dimensional state space, so a single linear map can send each trajectory to its own next motor command. The robot body's 20 triangular faces, arising from its equilibrium shape as Jessen's orthogonal icosahedron, provide a coarse partition of the state space that determines which attractor is reached.","core_discovery":"On the authors' own terms, the discovery is that a 48-dimensional reservoir state made of passive-tendon lengths and velocities is enough for a single linear readout, $W = DR^\\top(RR^\\top+\\beta I)^{-1}$, to reconstruct two different target motor signals once the loop is closed. The two outputs, a fast forward crawl (0.465 m/s, about 31 body lengths per minute) and an on-the-spot oscillation with one active actuator, are coexisting attractors of the closed-loop system. From the 20 initial states defined by each icosahedron face as the bottom face, the system converges to seven attractors in total: Attractor A from face 17, Attractor B from face 13, a fixed point from faces 3, 8, 15, 18, and 19, and periodic, quasi-periodic, or chaotic-looking attractors from the remaining faces. The bottom face is the main selector, and a perturbation that rolls the robot to a different face switches the behavior. The authors also map the regions of tendon stiffness where both trained attractors can be learned, and they classify bifurcations caused by parameter changes before training as pre-learning and after training as post-learning.","pith_inferences":["Editorial inference: if the bottom-face-to-attractor correspondence is a general property of this morphology, the number of faces of the tensegrity structure gives a natural upper bound on the number of coexisting behaviors, so body geometry directly sets behavioral capacity.","Editorial inference: the extreme final-state sensitivity near the basin boundary could be used as a deterministic physical random number generator or probabilistic behavior selector, since tiny changes in the driving parameters decide which attractor appears.","Editorial inference: the results suggest a content-addressable memory interpretation in which a short preview of a target behavior, not a full open-loop drive, might suffice to select the attractor; this could be tested by varying the preview length.","Editorial inference: post-learning bifurcations imply that gradual tendon aging or damage, without any retraining, could push the robot from one trained gait into new untrained gaits, turning wear into a potential source of behavioral diversity."],"forward_implications":["A single set of readout weights can control multiple qualitatively different robot behaviors, so behavior-specific controllers or additional readout modules are not needed for each gait.","Behavior selection can be achieved by changing the robot's initial body configuration, for example by nudging it onto a different bottom face, instead of by switching control parameters.","Untrained attractors are intrinsic outcomes of learning in an embodied system, so failed or unintended training results can be studied as a source of novel behaviors rather than discarded.","The region of tendon-stiffness parameters in which both trained attractors can be reconstructed defines a practical operating envelope for multifunctionality on this morphology.","Because the training procedure only needs separated trajectories and a linear readout, the approach transfers to other physical reservoirs, including photonic, spintronic, and quantum systems."],"supporting_citations":[{"why":"Defines multifunctional reservoir computing and provides the training-and-attractor-reconstruction method that this paper transfers to a physical tensegrity reservoir.","marker":"[22]"},{"why":"Introduces untrained attractors and basin analysis in a multifunctional reservoir, the framework the paper adopts for its attractor landscape.","marker":"[23]"},{"why":"Establishes physical reservoir computing as a field and justifies using material body dynamics as the computational substrate.","marker":"[30]"},{"why":"Shows tensegrity structures can act as physical reservoirs for locomotion control in simulation, the direct precursor this work extends.","marker":"[44]"},{"why":"Provides the hardware-validated tensegrity PRC implementation, supporting the practical plausibility of the simulated results.","marker":"[45]"},{"why":"Documents the repertoire of crawling, rolling, hopping, shaking, and chaotic behaviors of the same simulated tensegrity robot that this study draws on.","marker":"[56]"},{"why":"MuJoCo physics engine whose elastic-tendon and contact models generate all reported attractors, basins, and bifurcations.","marker":"[63]"},{"why":"Identifies the robot's equilibrium shape as Jessen's orthogonal icosahedron, which provides the 20-face labeling used as the behavior selector.","marker":"[64]"},{"why":"Supplies the fixed-random-reservoir, train-only-readout scheme underlying the linear readout used in this work.","marker":"[65]"},{"why":"NSGA-II genetic algorithm used to co-optimize the two Lissajous target signals for distance, behavioral difference, and learnability.","marker":"[93]"}],"fun_headline_variants":["Soft tensegrity robot computes seven behaviors with one readout","Single linear readout gives tensegrity robot multiple behaviors","Physical reservoir computing lets soft robot switch behaviors by rolling","Tensegrity robot uses body dynamics as reservoir for seven attractors","One readout, seven behaviors in a simulated tensegrity robot"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire result rests on the simulation: MuJoCo's soft-contact and elastic-tendon models must faithfully represent a real tensegrity robot's dynamics, because the attractors, basins, and bifurcations are all computed from simulated trajectories and no hardware validation is reported.","fun_headline_variants_meta":{"raw":{"variants":["Soft tensegrity robot computes seven behaviors with one readout","Single linear readout gives tensegrity robot multiple behaviors","Physical reservoir computing lets soft robot switch behaviors by rolling","Tensegrity robot uses body dynamics as reservoir for seven attractors","One readout, seven behaviors in a simulated tensegrity robot"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000263,"raw_usage":{"total_tokens":1618,"prompt_tokens":982,"completion_tokens":636,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":550}},"tokens_in":598,"tokens_out":636,"duration_ms":7231,"temperature":1.0,"reasoning_tokens":550,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:40:45.998937+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a physical six-bar tensegrity robot with the same geometry, tendon stiffness, damping, and actuator placement, run the identical open-loop training with the two Lissajous targets, and close the loop from the two final states. The central claim fails if the crawl and the on-the-spot oscillation do not reappear, or if the reported face-to-attractor map (for example, face 17 giving crawling and face 13 giving oscillation) is not reproducible across trials.","supporting_citations":[{"cited_title":"Caluwaerts , author J","cited_arxiv_id":null,"evidence_quote":"Provides the hardware-validated tensegrity PRC implementation, supporting the practical plausibility of the simulated results."},{"cited_title":"Terajima , author K","cited_arxiv_id":null,"evidence_quote":"Documents the repertoire of crawling, rolling, hopping, shaking, and chaotic behaviors of the same simulated tensegrity robot that this study draws on."},{"cited_title":"Todorov , author T","cited_arxiv_id":null,"evidence_quote":"MuJoCo physics engine whose elastic-tendon and contact models generate all reported attractors, basins, and bifurcations."},{"cited_title":"Gorkavyy \\ and\\ author D","cited_arxiv_id":null,"evidence_quote":"Identifies the robot's equilibrium shape as Jessen's orthogonal icosahedron, which provides the 20-face labeling used as the behavior selector."},{"cited_title":"echo state","cited_arxiv_id":null,"evidence_quote":"Supplies the fixed-random-reservoir, train-only-readout scheme underlying the linear readout used in this work."},{"cited_title":"Deb , author A","cited_arxiv_id":null,"evidence_quote":"NSGA-II genetic algorithm used to co-optimize the two Lissajous target signals for distance, behavioral difference, and learnability."}],"review_version":1}