{"id":"6d071a6b-77ec-4ce3-8d97-a21db7778571","arxiv_id":"2411.18423","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Evolving robot bodies with a homeokinetic controller as a cheap exploration proxy produces more diverse designs that score higher on downstream tasks than a fixed-controller baseline.","lead":"The authors replace the costly controller-training step in robot evolution with a homeokinetic controller, a simple self-organizing rule that makes almost any robot body explore on its own, and use that exploration as the evolutionary fitness. The resulting robot designs train into higher-scoring locomotion, manipulation, and hill-climbing robots than a fixed-controller baseline, using about 40 hours on 64 CPUs for 10,000 designs.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Asymmetric selection thresholds (0.4 vs 0.2) confound the MEHK-vs-MEFC downstream comparison; matched-threshold controls are missing.","rationale":"The paper makes a specific, falsifiable claim: MEHK (homeokinesis-based evaluation) produces designs that, after NCMA-ES training, outperform MEFC (fixed random controller) on four tasks. I read the experimental protocol carefully. The generation phase shows MEHK achieves far higher exploration scores than MEFC—expected, since homeokinesis adapts to the robot—so the exploration fitness distributions are strongly separated. The selection step in Section III-B then applies different absolute thresholds to the two methods (0.4 vs 0.2) and keeps three designs from a Pareto front. This is the point where the comparison becomes fragile. Because only three designs are trained per replicate (Section III-C), the identities of those designs determine the outcome. A higher threshold for MEHK means the MEHK pool is more stringently filtered; if the threshold for MEFC were raised, most MEFC designs would be discarded, but the authors chose a lower threshold to keep some MEFC designs. This is not an apples-to-apples comparison. The authors' own admission that exploration score has 'minimal impact' on downstream performance (Section IV-B) weakens the argument that the higher threshold is selecting for task-relevant viability. The observed morphological differences—MEHK designs have more wheels and fewer limbs (Fig. 6)—could be the real driver: wheeled robots are easier to train on locomotion and manipulation. This would mean the advantage is a byproduct of the evaluation policy's bias, not of a general viability proxy. A matched-threshold or quantile-based selection control would settle this. The paper has genuine strengths: the homeokinesis implementation is well described, the 30-replicate experiments are more thorough than typical, and the authors are candid about the proxy's limitations. However, the central comparison lacks the control needed to support the strong claim. The reader's verdict of CONDITIONAL is appropriate; the condition should be a matched-threshold sensitivity analysis.","tokens_in":9424,"tokens_out":4627,"duration_ms":40827,"concrete_test":"Run the full pipeline (generation with MEHK and MEFC, selection, NCMA-ES training) under matched selection thresholds: (i) threshold 0.4 for both, (ii) threshold 0.2 for both, and (iii) equal quantile thresholds (e.g., top 20% of each method's exploration score distribution). For each setting, train 3 selected designs per replicate over 30 replicates and compare downstream scores on all four tasks. Also report the number of designs passing each threshold and the morphological composition (wheels, limbs, sensors) of the selected sets. If MEHK's advantage diminishes or reverses under any matched threshold, the central claim is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—MEHK designs outperform MEFC designs after NCMA-ES training—is established through a downstream-selection step with asymmetric thresholds. Section III-B states: \"a selection is made by keeping the designs with a fitness value above a set threshold (0.4 for MEHK and 0.2 for MEFC).\" Only three designs per replicate are then trained (Section III-C). Because MEHK's exploration scores are systematically higher (Fig. 5), a 0.4 threshold is less restrictive for MEHK than a 0.2 threshold is for MEFC, and the two selected sets are not drawn from comparable regions of the fitness distribution. The authors acknowledge in Section IV-B that exploration score has \"minimal impact\" on downstream task performance, so the higher threshold does not obviously select for trainability. The asymmetry could instead select for morphological traits (e.g., more wheels, fewer limbs; Fig. 6) that make the downstream tasks easier, independent of homeokinesis. No threshold sensitivity analysis or matched-threshold control is reported, so the headline superiority may be a selection artifact.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes MEHK, a morpho-evolution algorithm that evaluates robot designs using a homeokinetic controller to generate exploratory behavior instead of training a task-specific controller during evolution. The authors compare MEHK to MEFC, a morpho-evolution baseline with a fixed feed-forward controller, on four downstream tasks after selecting three designs per replicate via a fitness threshold and a Pareto-based sparsity selection. They report that MEHK-generated designs achieve higher task scores on all four tasks, similar wall-clock time (about 40 hours for 10,000 designs on 64 CPUs), and different morphological diversity compared to MEFC. The manuscript includes 30 replicates per condition and reports Brunner-Munzel tests with p-values below 1e-4 on most tasks.","tokens_in":9698,"tokens_out":5038,"duration_ms":46595,"significance":"If the result holds, MEHK offers a computationally cheap method for generating robot morphologies without per-design learning, potentially scaling to larger design spaces. The paper is, to my knowledge, the first to combine homeokinesis with CPPN-based morpho-evolution and to include exteroceptive sensors in this framework. Strengths include the clean 30-replicate protocol, the use of statistical testing, and the analysis of morphological distributions. However, the central comparison is burdened by asymmetric selection thresholds between MEHK and MEFC, and the claimed comparison with 'current MEL methods' is not actually performed; the only baseline is a fixed-controller evolution. These issues require careful re-analysis and re-framing before the headline claims can be accepted.","major_comments":[{"comment":"The selection step uses asymmetric thresholds: '0.4 for MEHK and 0.2 for MEFC.' Since Fig. 5 shows MEHK exploration scores are systematically higher (median 0.8 vs 0.45), these thresholds select different percentiles of the two distributions. The paper's own Section IV-B states that exploration score has 'minimal impact' on downstream performance, which means the higher threshold does not have an obvious mechanism for selecting more trainable designs. The observed downstream advantage could therefore be caused by the threshold asymmetry or by morphological correlates (e.g., more wheels and fewer limbs in MEHK, per Fig. 6) rather than by homeokinesis acting as a viability filter. The authors should provide a threshold-sensitivity analysis with matched thresholds, or compare downstream performance of all generated designs without threshold filtering.","section":"Section III-B (Robotic design selection)"},{"comment":"The abstract claims comparison with 'current MEL methods,' but the only in-paper baseline is MEFC, a morpho-evolution with a fixed controller and no learning phase (Section II-A-c). This overstates the evidence: the efficiency argument relies on a comparison with prior work (Gupta et al. [6]) rather than an in-paper MEL baseline. Moreover, Fig. 5 shows MEHK and MEFC use the same wall-clock time, so the abstract's 'quickly generated compared to morpho-evolution with static parameters' is not supported by the reported experiment. Please either add a genuine MEL baseline (e.g., AME with CMA-ES evaluation) or carefully revise the claims and abstract to state that the comparison is against fixed-controller morpho-evolution.","section":"Abstract and Section I (Introduction)"},{"comment":"The paper motivates homeokinesis as a proxy for design viability: 'homeokinesis is a proxy towards optimal designs, filtering designs unsuitable for movement' (Section I) and 'The exploration task biases the designs for functional combinations...' (Section III-A). Yet Section IV-B states that 'the exploration score... has minimal impact on the performance of the downstream tasks.' The mechanism by which MEHK improves downstream performance is therefore not established, and the downstream advantage could be a side-effect of selection on morphological traits. I recommend reporting the correlation between exploration fitness and downstream task scores within each condition, and comparing random subsamples with matched thresholds, to directly test the proxy-validity claim.","section":"Section I and Section IV-B"}],"minor_comments":[{"comment":"The caption contains a typo: 'homoekinesis' should be 'homeokinesis.'","section":"Figure 2 caption"},{"comment":"The word 'Unver' should likely be 'University' in 'Edinburgh Napier Unver.'","section":"Acknowledgements"},{"comment":"The sparsity score uses 15 nearest neighbours; no justification or sensitivity analysis is provided for this choice.","section":"Section III-B"},{"comment":"The hill-climbing advantage is modest (medians around 0.4 vs 0.5) and only a tail of MEHK robots performs substantially better; reporting effect sizes or confidence intervals alongside p-values would strengthen the conclusions.","section":"Section IV-B"},{"comment":"The statement that MEHK 'can generate a wider diversity of designs than MEFC when comparing components' needs reconciliation with Fig. 7, which shows MEFC has higher chassis diversity; the claims about diversity are therefore more nuanced than the text suggests.","section":"Section IV-A"},{"comment":"The paper states 'The code will be available upon acceptance' but no code or detailed simulation parameters (physics engine, timestep, sensor noise model) are provided; for a computational paper, shipping code or a complete parameter listing would substantially improve reproducibility.","section":"Section III-C and reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The gap between the abstract's 'current MEL methods' claim and the actual fixed-controller baseline is significant, and the asymmetric selection thresholds are a potential confound that undermines the headline superiority result. This is fixable but requires new experiments (matched thresholds, possibly a true MEL baseline) rather than modest revisions. As editor, you may also wish to consider whether the journal's scope is well served by a paper that does not yet include a direct MEL comparison despite its title and abstract; I would recommend requesting a revision that either adds such a baseline or explicitly restricts claims. Please also encourage the authors to release code, as the current 'available upon acceptance' statement is not verifiable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe genuinely new piece here is using homeokinesis as a fast, task-agnostic evaluation proxy during morpho-evolution, including exteroceptive sensors. The MEHK vs MEFC comparison is replicated 30 times, and the exploration-phase results look solid: homeokinesis covers more of the arena and produces different morphological biases than a fixed random network. The paper also makes a useful practical point: 10,000 designs in roughly 40 hours on 64 CPUs is much cheaper than deep-RL MEL, which the authors fairly attribute to prior work.\n\nThe soft spots sit in the downstream comparison. The abstract says \"compared with current MEL methods,\" but the experiments only include MEFC, a fixed-controller baseline; no learning-based MEL baseline appears. More importantly, the selection step before NCMA-ES training uses different fitness thresholds for the two methods: 0.4 for MEHK and 0.2 for MEFC. Since MEHK's exploration scores are systematically higher, the 0.4 cut is less restrictive and the two selected sets are not drawn from comparable regions of the fitness distribution. The authors themselves note in Section IV-B that exploration score has minimal impact on downstream performance, which undercuts the idea that the higher threshold selects for trainability. The asymmetry could instead favor morphological traits, such as the wheel-heavy designs shown in Fig. 6, that make the downstream tasks easier. There is no matched-threshold control and no threshold sensitivity analysis. That is a real confound, not a nitpick.\n\nTo the authors' credit, the proxy assumption itself is honestly discussed: exploration is necessary but not sufficient, and the paper says so. I do not think that alone sinks it, but it does mean the headline claim that MEHK designs \"score higher in all the tasks\" is not yet established as a property of homeokinesis per se. The missing code and data also make it hard to verify the cost and reproducibility claims.\n\nWho is this for? Evolutionary robotics researchers, especially those working on morpho-evolution, will want to read it. It deserves a serious referee: the method is clearly described, the core comparison is replicated, and the threshold confound is addressable with additional controls. I would not cite the downstream superiority claim as established, but I might cite the approach as a cost-reducing alternative worth testing.","headline":"Homeokinesis as a cheap evaluation proxy in morpho-evolution is a real idea, but the downstream superiority claim is confounded by asymmetric selection thresholds and a missing MEL baseline.","tokens_in":10180,"tokens_out":2002,"would_cite":true,"duration_ms":18795,"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":"Homeokinesis makes evolved robot designs outperform fixed-controller baselines on every tested task.","keywords":["morpho-evolution","homeokinesis","intrinsic motivation","generative robot design","evolutionary algorithms","task-agnostic design","diversity"],"falsifier":"Evolve designs with MEHK but replace the homeokinetic score with a random score that matches the same distribution of exploration values, keeping the Pareto selection and NCMA-ES training identical; if downstream performance does not drop, the homeokinetic signal itself is not carrying the result.","tokens_in":9215,"feed_emoji":"🤖","tokens_out":4592,"duration_ms":39091,"temperature":0.7,"pith_summary":"The paper claims that replacing the expensive controller-learning phase in evolutionary robot design with a homeokinetic controller—an intrinsic-motivation rule that makes any embodied robot explore its surroundings within seconds—lets evolution generate robot bodies faster, in greater diversity, and with better downstream performance. The authors compare their method, MEHK, against morpho-evolution with a fixed random controller (MEFC) on four tasks: hill climbing, flat and rough locomotion, and object manipulation. MEHK-generated designs score higher in all four tasks, reach higher exploration scores during generation, and produce a more even spread of component types, using the same wall-clock time as the baseline. The practical stake is that co-evolving morphology and control can scale to larger design spaces without the compute cost that has blocked prior morpho-evolution-with-learning systems.","feed_headline":"Homeokinesis speeds robot design evolution and improves all four tasks","feed_subtitle":"A short intrinsic-motivation phase replaces costly learning, generating 10,000 diverse bodies in about 40 hours.","key_machinery":"The load-bearing object is the homeokinetic controller, a pseudo-linear controller updated by stochastic gradient descent on the time-loop error $\\mathrm{TLE} = \\| L^{-1} E \\|^2$, where $E$ is the forward-model prediction error and $L$ the Jacobian of the sensorimotor loop. Minimising the TLE balances predictable behaviour against sensitive behaviour, producing exploratory motion in a few seconds with no task reward. In MEHK this controller replaces the learning phase: each design's fitness is the fraction of 64 arena cells visited during 20 minutes of homeokinetic exploration, and evolution runs through the asynchronous AME algorithm on CPPN-encoded bodies.","core_discovery":"On the paper's own terms, the central discovery is that homeokinesis, a parameter-free intrinsic-motivation controller that maximises predictability and sensitivity via the time-loop error, is an effective and cheap evaluation policy for morpho-evolution. By scoring each robot body on how much of an 8-by-8 arena it covers in 20 minutes of homeokinetic exploration, the evolutionary loop can filter for functionally viable bodies—ones that move and interact with the environment—without task-specific training. From 10,000 generated designs, the authors select three per replicate from a Pareto front of exploration fitness and morphological sparsity, train Elman-network controllers with NCMA-ES, and find that the MEHK bodies outperform MEFC bodies on all four downstream tasks, with statistical significance at p<$10^{-4}$. The method generates 10,000 robots in about 40 hours on 64 CPUs, which the authors contrast with the 1,152 CPUs and 4,000 designs reported for a deep-RL morpho-evolution baseline.","pith_inferences":["The paper's own observation that the exploration score has minimal impact on downstream performance suggests the advantage of MEHK may come primarily from the diversity of body morphologies it produces rather than from homeokinesis selecting intrinsically 'good' bodies; a selection mechanism preserving the same diversity might perform equally well.","The same morpho-evolution framework could serve as a testbed for comparing other intrinsic-motivation signals, such as empowerment or predictive information, as cheap viability filters for robot bodies.","The fixed-controller baseline's random weights bias it toward jointed bodies, so part of MEFC's weakness may stem from a controller-design mismatch rather than the absence of intrinsic motivation; a stronger baseline would use a controller tuned per morphology without full task-specific learning."],"forward_implications":["MEHK designs outperform MEFC designs in all four downstream tasks, so the homeokinetic exploration score is a viable cheap proxy for body viability in morpho-evolution.","The method generates and evaluates 10,000 robot designs in about 40 hours on 64 CPUs, substantially lowering the compute barrier compared to deep-RL morpho-evolution approaches.","MEHK produces a more diverse mixture of component types than MEFC, countering the premature convergence that learning-based evaluation tends to induce.","The Pareto selection between exploration fitness and morphological sparsity yields both generalist robots and specialists that excel at single tasks."],"supporting_citations":[{"why":"Supplies the homeokinetic controller and time-loop error learning rule that forms the evaluation policy.","marker":"[26]"},{"why":"Supplies the asynchronous morpho-evolution (AME) algorithm used as the evolutionary backbone.","marker":"[27]"},{"why":"Supplies the CPPN encoding that maps spatial coordinates to robot body components.","marker":"[28]"},{"why":"Supplies the NCMA-ES algorithm with novelty search used to train downstream task controllers.","marker":"[29]"},{"why":"Provides the deep-RL morpho-evolution comparison the authors cite for computational cost and design count.","marker":"[6]"},{"why":"Provides the comparison showing classic morpho-evolution runs far more generations than MEL under equal compute.","marker":"[17]"},{"why":"Supplies the closest prior task-agnostic morphology evolution method using empowerment as a proxy.","marker":"[24]"},{"why":"Supplies the design-space definition with five components and the voxel-based chassis.","marker":"[30]"}],"fun_headline_variants":["Homeokinesis makes robot evolution faster and more diverse","Intrinsic motivation cuts design time and boosts task scores","Homeokinetic exploration yields diverse robots that excel on all tasks","10,000 diverse robot bodies in 40 hours via intrinsic motivation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 20-minute exploration score under homeokinesis is a valid proxy for whether a robot body will train well on unrelated later tasks; the authors themselves note the score has minimal impact on downstream performance.","fun_headline_variants_meta":{"raw":{"variants":["Homeokinesis makes robot evolution faster and more diverse","Intrinsic motivation cuts design time and boosts task scores","Homeokinetic exploration yields diverse robots that excel on all tasks","10,000 diverse robot bodies in 40 hours via intrinsic motivation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000719,"raw_usage":{"total_tokens":3236,"prompt_tokens":960,"completion_tokens":2276,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":2208}},"tokens_in":576,"tokens_out":2276,"duration_ms":13727,"temperature":1.0,"reasoning_tokens":2208,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:12:49.270923+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evolve designs with MEHK but replace the homeokinetic score with a random score that matches the same distribution of exploration values, keeping the Pareto selection and NCMA-ES training identical; if downstream performance does not drop, the homeokinetic signal itself is not carrying the result.","supporting_citations":[{"cited_title":"Der and G","cited_arxiv_id":null,"evidence_quote":"Supplies the homeokinetic controller and time-loop error learning rule that forms the evaluation policy."},{"cited_title":"Improving efficiency of evolving robot designs via self-adaptive learning cycles and an asynchronous archi- tecture,","cited_arxiv_id":null,"evidence_quote":"Supplies the asynchronous morpho-evolution (AME) algorithm used as the evolutionary backbone."},{"cited_title":"Compositional pattern producing networks: A novel abstraction of development,","cited_arxiv_id":null,"evidence_quote":"Supplies the CPPN encoding that maps spatial coordinates to robot body components."},{"cited_title":"Sample and time efficient policy learning with cma-es and bayesian optimisation,","cited_arxiv_id":null,"evidence_quote":"Supplies the NCMA-ES algorithm with novelty search used to train downstream task controllers."},{"cited_title":"The effects of learning in morphologically evolving robot systems,","cited_arxiv_id":null,"evidence_quote":"Provides the comparison showing classic morpho-evolution runs far more generations than MEL under equal compute."},{"cited_title":"Task-agnostic morphology evo- lution,","cited_arxiv_id":null,"evidence_quote":"Supplies the closest prior task-agnostic morphology evolution method using empowerment as a proxy."},{"cited_title":"Bootstrap- ping artificial evolution to design robots for autonomous fabrication,","cited_arxiv_id":null,"evidence_quote":"Supplies the design-space definition with five components and the voxel-based chassis."}],"review_version":1}