{"id":"474806e2-6008-4007-86d4-6982bf3c184d","arxiv_id":"2505.12029","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A hexapod using GOLLUM autonomously learned multiple locomotion skills in about an hour by growing new neural subnetworks for new conditions and by blending previously learned skills through supplementary connections.","lead":"GOLLUM is a control network that lets a physical six-legged robot learn new walking gaits by adding small neural modules when it meets a new terrain, and by reusing older gaits as starting points. The robot learned several skills within about an hour on real hardware, with no simulation pretraining and no task context from a human.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fixed four-action triangular-basis CPG restricts motor trajectories to a low-dimensional subspace, so the asserted 'full action space' and the 'no gait parameterization' part of the central claim are not secured.","rationale":"The reader's conditional verdict already flags the hand-set basis parameters; my analysis agrees and sharpens it. The key question is not whether the reported hexapod skills were learned (they appear to be, with videos, code, and ablations), but whether the framework supports the broad claim of full-action-space, parameterization-free lifelong learning. Eqs. 5-7 show a representational restriction: all motor outputs lie in the span of a small fixed set of triangular basis functions per active subnetwork. This is a correctness risk for the 'only approach' claim because if the rhythmic prior is doing the work, the comparison against methods with different priors is not apples-to-apples, and the 'without human intervention' wording is too strong. The concern is testable by a projection or approximation experiment, and if the residual is small for all gaits of interest, the paper's claim survives; if not, the claim should be narrowed to the demonstrated gait family and the hand-designed rhythmic prior should be reported as a design constraint rather than as absence of parameterization. I therefore keep the reader's conditional verdict rather than escalating to rejection, since the empirical demonstrations and provided code are real evidence that the core mechanism works within the demonstrated regime.","tokens_in":29813,"tokens_out":10466,"duration_ms":118623,"concrete_test":"Construct a target 18xT trajectory for a hexapod wave gait whose six leg-lift phases are spaced at 0, 1/6, 2/6, 3/6, 4/6, and 5/6 of the gait cycle. Using the released GOLLUM code, generate the B_l[t] basis signals from Eqs. 5-7 with the paper's tau settings and compute, for each motor, the least-squares projection M_j[t] = sum_l c_{j,l} B_l[t] onto the span of those bases. If the normalized L2 residual exceeds 10% of the target signal amplitude, the full-action-space assertion fails. As a control, repeat the same projection for the tripod and bouncing gaits actually learned; if those residuals are near zero while the wave-gait residual is large, the architecture is confirmed to encode a specific rhythmic prior rather than a full action space.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim says GOLLUM learns diverse new skills with no gait parameterization and no human intervention. That claim leans on the assertion, in the Sequential Central Pattern Generator section, that GOLLUM incorporates a rhythmic prior while maintaining a full action space. The equations do not support this. Combining Eqs. 5-7, PM_i[t] sums B_k[t] through fixed or learned weights and M_j[t] sums PM_k[t], so every motor command is a linear combination of the triangular basis signals B_k[t]. With four C neurons and four actions per subnetwork, a single active subnetwork provides at most four temporal basis functions; the off-diagonal wPMi,Bk only recombines these same bases and cannot create new temporal shapes. The set of representable 18-motor trajectories over a gait cycle is therefore a low-dimensional subspace, not the full action space. A skill whose optimal gait needs more than four distinct phase offsets per cycle (e.g., a hexapod wave gait with six individually timed leg lifts) cannot be represented. Moreover, the B-layer off-diagonal weights are selected empirically and the C-layer parameters are set analytically in a supplementary document with unresolved equation and figure references in the arXiv text, so the rhythmic prior is a hand-designed component. If this hand-designed prior is the real source of sample efficiency, the claims of autonomy and of no gait parameterization are overstated. The reported successes on tripod-like and bouncing gaits do not exercise the full action space and therefore do not resolve the concern.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces GOLLUM, a modular interpretable neural controller for legged robots that grows new CPG-like ring subnetworks when online-fitted value and observation prediction boundaries are exceeded. Primary learning updates PM-to-M weights to shape motor commands for the active skill, while supplementary learning updates B-to-PM weights to recombine previously learned action patterns. Experiments on the MORF hexapod report flat-ground walking from scratch in about 200 episodes, continual acquisition of slope, deformable-terrain, and motor-dysfunction skills within about an hour, recall of old skills, and ablations suggesting that neurogenesis and supplementary learning prevent catastrophic forgetting and exploit similarity. The paper claims that GOLLUM is the only real-world locomotion learning framework that achieves online continual learning without pretraining or task context.","tokens_in":30138,"tokens_out":6549,"duration_ms":67753,"significance":"If the empirical claims hold, GOLLUM is a valuable contribution: it demonstrates that a strongly structured, interpretable controller can be grown online on physical hardware, with substantially better sample efficiency than the reported deep RL baselines, and it provides a concrete mechanism—separate primary and supplementary weight spaces—by which new skills can be added without interfering with old ones. The provision of videos and open code is a strength, as are the quantitative compactness and fidelity comparisons. The significance is currently capped by two gaps: the lack of a formal characterization of the representable motor-command space, and the under-specified statistics for the continual-learning ablations; both need to be fixed before the strongest claims can be accepted.","major_comments":[{"comment":"The claim in this section that GOLLUM 'incorporates a rhythmic prior ... while maintaining a full action space' is not supported by the equations. From Eqs. (5)–(7), every motor command is a linear combination of the triangular basis signals B_k[t], because PM is a linear map of B and M is a linear map of PM. With four C neurons and four B bases per subnetwork, a single active subnetwork spans at most four temporal basis functions, and the off-diagonal weights w_{PM_i,B_k} merely recombine those same bases without enlarging the temporal shape space. The paper therefore has not established that arbitrary leg-coordination patterns (e.g., a hexapod wave gait with six independent phase offsets) can be represented. Please either prove the coverage claim for the relevant gait class, or replace 'full action space' with a precise statement of the low-dimensional subspace that the architecture can represent and explain why that subspace suffices for the demonstrated skills.","section":"Sequential Central Pattern Generator (C), Eqs. (5)–(7)"},{"comment":"The basis-shape parameters are central to the representation, but their specification is incomplete. The text states that the C-layer parameters are 'set analytically as described in Eqs.??–?? (also in Figures?? and??) in the supplementary material,' and the off-diagonal B-layer weights are 'selected empirically as depicted in the supplementary document.' The unresolved equation and figure numbers and the empirical selection leave the actual basis functions unspecified in the manuscript. Because Eqs. (5)–(7) show that all learned motor behavior is a linear combination of these fixed bases, the reader cannot reproduce the approach or verify the claimed absence of gait parameterization without these parameter values and their selection procedure.","section":"Basis (B), Eq. (5)"},{"comment":"The continual-learning support for the main claims rests on paired t-tests with n=20, but the manuscript does not define what n is. If n refers to episodes within a single physical run, those comparisons are pseudoreplicated and cannot support the stated conclusions; if n refers to independent runs, the number of runs and the per-condition variance need to be reported. The same issue affects the speed comparisons in 'Primitive Locomotion Learning' (n=20, Figure 8). Please state the experimental design, report means and error bars for repeated runs, and adjust the statistical claims accordingly.","section":"Experiments and Results, statistical tests in Figures 11–14"},{"comment":"The novelty detector is itself adaptive: Vδ and Oδ are trained online to cover the observed reward and feedback deviations, and the text reports that a reward-only drop at episode 100 expanded the prediction boundary without triggering neurogenesis. This creates a need to specify the condition under which boundary expansion stops and a new subnetwork is created; otherwise the mechanism is indistinguishable from an always-expanding threshold, and the claimed autonomy of 'detecting new conditions' is not established. A worked analysis of the boundary dynamics at the episode-100 versus episode-150 transitions would resolve this concern.","section":"Subnetwork Neurogenesis, Eqs. (8)–(11)"},{"comment":"The statement that GOLLUM is 'the only locomotion learning framework' with the listed properties is stronger than the evidence in Table 2. The table covers 18 selected methods and gives no systematic inclusion criteria or search protocol, and several table entries are not directly comparable because they differ in robot platform, simulation-versus-real setting, and task definition. Please narrow the claim to the methods actually compared, or provide a systematic literature selection protocol that justifies 'only.'","section":"Discussion, final paragraph of 'Life-long Locomotion Learning Research Aspect'"}],"minor_comments":[{"comment":"The term w_{Ci,Ii} I_i[t] is inconsistent with the text's description that the selection input is I_{i+1}; align the index.","section":"Eq. (4)"},{"comment":"The relation between τ_i in the I-layer and w_{Bi,Bi}=1−τ_i in the B-layer should be stated explicitly; as written, 'set to the same value as that in the basis layer' is ambiguous.","section":"Eqs. (3) and (5)"},{"comment":"There are several typos: 'nework' in the Output Layer subsection, 'within an hours' in General Continual Locomotion Learning, 'refered to behavior model' and 'refereed to' in Separation and Incrementation, and 'unsupervisely' in the Abstract.","section":"Throughout"},{"comment":"The notation t+H min τ=t r[τ] is malformed; use min_{t≤τ≤t+H} r[τ].","section":"Eq. (19)"},{"comment":"Figure 15a is labeled a conceptual representation; consider stating in the caption that the axes are illustrative, since the figure is used to support the interpretability/performance trade-off claim.","section":"Figure 15a"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is likely within scope for a robotics or neural-control journal. The 'only' claim will need editorial scrutiny, and the GitHub repository was not reviewed in this report; if the repository is part of the submission, it should be checked to confirm that it contains the parameter tables referenced by the unresolved Eqs.?? and Figures?? in the text."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this is a genuine physical-robot continual learning story, not a simulation-only one. A hexapod learns several gaits in about an hour with no pretraining, and the ablation plus videos make a plausible case that neurogenesis and the dual-layer learning each contribute. It deserves a serious referee. But the paper's strongest claim—that the rhythmic prior retains a full action space—is not supported by the equations, and the authors should be asked to either prove it or drop it.\n\nWhat is actually new is the architecture itself: growable CPG columns created by unsupervised neurogenesis driven by prediction-error, plus a supplementary learning layer that recombines previously learned skills. That combination is not in the authors' prior VMNC work or in the cited literature. The physical experiments on slopes, soft terrain, and a frozen joint are the right demonstrations for the forgetting and transfer claims, and the interpretability metrics are a step beyond the usual hand-waving.\n\nThe soft spot is the full-action-space assertion. From Eqs. 5–7, each motor command is a linear combination of at most four triangular basis signals per active subnetwork. The off-diagonal PM weights recombine those same bases; they do not create new temporal shapes. So the representable movements are a low-dimensional subspace, not an 18-dimensional action space. The tested gaits (tripod-like, bouncing) live in that subspace, so the experiments do not exercise the gap. This does not sink the empirical contribution—the inductive bias is probably why learning is so fast—but it does mean the paper overstates how unconstrained the learning is.\n\nThe statistical reporting is thin in places: n=20 paired t-tests without saying whether these are episodes or independent runs, and the continual learning curves are single trajectories with no error bars. The supplementary material is referenced with unresolved equation numbers in the arXiv text, so the analytical parameter settings cannot be checked from the paper. These are fixable but should be fixed.\n\nWho is the paper for? People who care about real-world continual learning for legged robots and interpretable controllers. It is not a general continual RL theory paper. If the authors revise to hedge the full-action-space claim and supply the missing details, it becomes a solid systems and empirics paper.\n\nRecommendation: send to peer review. The core mechanism is clear and the physical results are worth referee time. A good referee should push on the representational claim, not on whether the robot learned anything.","headline":"A real-robot continual learning result with real strengths, but the 'full action space' claim does not survive contact with its own equations; worth peer review, not worth taking at face value.","tokens_in":30666,"tokens_out":2925,"would_cite":true,"duration_ms":30331,"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":"GOLLUM claims a real hexapod can learn new gaits continuously, keep old ones, and reuse them, all without pretraining or task context.","keywords":["continual learning","locomotion learning","central pattern generator","interpretable neural control","neurogenesis","reinforcement learning","hexapod robot","catastrophic forgetting"],"falsifier":"Run GOLLUM on a terrain whose optimal gait requires an inter-leg phase pattern that is not generated by the four-action ring sequence, for example a prescribed asymmetric tripod. If the controller either fails to reach the reward level achieved by a directly parameterized gait controller, or needs neurogenesis to keep adding columns without converging, the assumption that the fixed triangular bases span a full action space is falsified.","tokens_in":29613,"feed_emoji":"🤖","tokens_out":9835,"duration_ms":95241,"temperature":0.7,"pith_summary":"GOLLUM (Growable Online Locomotion Learning Under Multicondition) is a control-and-learning framework that lets a real hexapod robot acquire new locomotion skills while it is running, keep the skills it has already learned, and reuse those skills to make later learning faster. The paper's central claim is that interpretability, not more capacity or more data, is what resolves the four standard problems of lifelong locomotion learning: sample inefficiency, lack of knowledge reuse, catastrophic forgetting, and incomprehensibility. On a physical hexapod, GOLLUM learned walking, slope climbing, and gaits for deformable terrains from scratch within roughly 100 to 200 episodes, about ten to twenty minutes per skill, using a single simple reward term. It later combined previously learned skills to improve performance in new conditions, and the paper claims this makes GOLLUM, among the compared approaches, the only one that realizes online continual locomotion learning in the real world without task context or human intervention in the tested setting.","feed_headline":"Robot learns new gaits in minutes and keeps old ones","feed_subtitle":"GOLLUM's interpretable ring network adds skills online without pretraining or task context, in 10–20 minutes per skill.","key_machinery":"The load-bearing object is the interpretable neural control network, a discrete-time non-spiking network organized in two dimensions at once. Horizontally, each column is a ring-like central pattern generator (CPG) subnetwork, a loop of neurons that keeps producing rhythmic patterns without external input, and each column is meant to encode one primary skill. Vertically, seven layers form four functional modules: feedback and two preprocessing layers ($FB$, $I'$, $I$) classify and gate the current condition; a sequential central pattern generator layer ($C$) emits discrete internal states; a basis layer ($B$) smooths them into sparse triangular signals; a premotor layer ($PM$) holds shared action patterns; and an output layer ($M$, $V$, $O$) maps patterns to motor commands, value predictions, and observation predictions. Three mechanisms carry the argument: primary learning updates only the active column's $PM\\to M$ weights, so inactive skills stay untouched; supplementary learning adjusts the $B\\to PM$ weights that blend skills; and neurogenesis, triggered when the return falls below the value-prediction boundary while an observation exceeds its prediction boundary, appends a column initialized from the most similar existing skill. The claimed full action space is preserved because the ring structure imposes a rhythmic prior while the bases are combined linearly, and the claimed interpretability follows from the sparse structure, where the connection matrix $\\kappa$ displays which behaviors transition to which.","core_discovery":"The central claim is that lifelong locomotion learning can be solved by making the controller itself interpretable rather than by adding complicated memory or regularization mechanisms. On the paper's own terms, GOLLUM's contribution is the combination of column-wise skill encoding, layer-wise function encoding, a dual-layer learning rule, and neurogenesis-driven network growth. The experiments report that this combination lets a physical hexapod learn walking from scratch on level ground in about 200 episodes, roughly ten minutes, acquire separate skills for slopes, a frozen joint, and soft or rough terrains in about an hour, recall earlier skills after later training without performance loss, and improve a recalled level-ground skill by about 40 percent by mixing in slope skills. The Discussion then states the strongest form of the claim: GOLLUM is the only locomotion learning framework that achieves online continual locomotion learning in the real world without catastrophic forgetting while exploiting task similarity during both direct knowledge transfer and learning, and without task context or human intervention, under unlimited space.","pith_inferences":["The dual-layer separation, per-skill primary weights plus a small mixing layer, is arguably a general recipe for lifelong learning: it avoids forgetting by construction rather than by regularization or replay, so it could be ported to other control problems such as manipulation or navigation wherever an interpretable rhythmic prior is available.","The prediction-boundary trigger for neurogenesis is an intrinsic novelty signal; a natural extension is to compare GOLLUM against common intrinsic-motivation bonuses such as prediction-error or count-based exploration, and to test whether the explicit two-channel surprise, value plus observation, is necessary or whether either channel alone suffices.","The strongest claim is explicitly qualified as holding under unlimited space; a stress test that would bound the claim is to check how skill count, memory, and learning time scale as the number of conditions grows, since the paper demonstrates five to six skills within an hour but does not establish an unbounded regime.","Because the basis-layer parameters are partly selected empirically, one could ask whether learning those parameters online, while keeping the sparsity structure, improves generalization to terrains with very different gait dynamics, or whether it breaks the layer-wise interpretability the approach relies on."],"forward_implications":["A legged robot can be deployed without simulation pretraining and still acquire new skills on the fly: GOLLUM reports usable gaits from scratch in roughly 100 to 200 episodes, about ten to twenty minutes per condition on the real robot.","Skills learned earlier remain recallable after later training: when the robot returned to a level floor after climbing a 25-degree slope, it recovered its earlier flat-ground gait instead of the sliding-backward behavior learned on the slope.","New conditions can be learned faster by reusing old skills: initializing a new column from the most similar prior skill, plus supplementary blending, let the robot combine a slope skill and a motor-dysfunction skill to climb a slope with a frozen joint in about 30 episodes.","Because the learned controller is interpretable, a designer can read the behavior hierarchy from the connection matrix and modify parameters such as the frequency parameter without retraining, and post-hoc decision-tree explanations of GOLLUM are reported to be smaller and lower-error than those for the compared deep networks.","The same core network is claimed to transfer to other embodiments and tasks: changing the output dimension from 18 to 12 joints adapts it to a quadruped, and replacing the locomotion reward with a fitting error gives a programming-by-demonstration method."],"supporting_citations":[{"why":"Supplies the CPG/RBF baseline (CPGRBF + PIBB) and the episodic replay schedule used in GOLLUM's learning loop.","marker":"Thor et al. (2020)"},{"why":"Provides the versatile modular locomotion controller that GOLLUM contrasts with, since its indirect encoding and pretraining are replaced by direct real-world learning.","marker":"Thor and Manoonpong (2022)"},{"why":"Gives the ring-network CPG oscillator construction on which the sequential central pattern generator layer is based.","marker":"Pasemann et al. (2003)"},{"why":"Supplies the parameter-exploring policy gradient rule that GOLLUM modifies with sparse masks and gradient weighting.","marker":"Sehnke et al. (2010)"},{"why":"Contributes the episodic policy-improvement perspective underlying the dual-layer weight updates.","marker":"Stulp and Sigaud (2012)"},{"why":"Provides the neural-visualization tool used to display and verify column-wise and layer-wise activities.","marker":"Srisuchinnawong et al. (2021a)"},{"why":"Supplies the compactness and completeness metrics used in the quantitative interpretability comparisons.","marker":"Nauta et al. (2023)"},{"why":"Defines the MORF hexapod robot used in all experiments.","marker":"Thor et al. (2018)"},{"why":"Motivates the use of value and observation prediction boundaries as surprise and uncertainty signals for neurogenesis.","marker":"Angela and Dayan (2005)"},{"why":"Provides the low-pass filter dynamics used to select the basis-layer parameters.","marker":"Srisuchinnawong et al. (2023)"}],"fun_headline_variants":["Hexapod robot teaches itself new skills, remembers old ones","Interpretable neural net lets robot keep learning without forgetting","Robot's growing brain learns new gaits while retaining old","Online continual learning: robot adds skills, never forgets","Growable neural control enables lifelong robot learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme depends on the idea that the hand-built ring network, with four action phases per skill and manually chosen smoothing parameters, can still produce every movement pattern the robot might need. If a useful gait cannot be written as a weighted sum of those fixed triangular bases, then the promise of learning diverse new skills without re-parameterizing the gait fails.","fun_headline_variants_meta":{"raw":{"variants":["Hexapod robot teaches itself new skills, remembers old ones","Interpretable neural net lets robot keep learning without forgetting","Robot's growing brain learns new gaits while retaining old","Online continual learning: robot adds skills, never forgets","Growable neural control enables lifelong robot learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000272,"raw_usage":{"total_tokens":1657,"prompt_tokens":996,"completion_tokens":661,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":582}},"tokens_in":612,"tokens_out":661,"duration_ms":6692,"temperature":1.0,"reasoning_tokens":582,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:42:26.534918+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GOLLUM on a terrain whose optimal gait requires an inter-leg phase pattern that is not generated by the four-action ring sequence, for example a prescribed asymmetric tripod. If the controller either fails to reach the reward level achieved by a directly parameterized gait controller, or needs neurogenesis to keep adding columns without converging, the assumption that the fixed triangular bases span a full action space is falsified.","supporting_citations":[{"cited_title":"IEEE transactions on neural networks and learning systems 32(9): 4013--4025","cited_arxiv_id":null,"evidence_quote":"Supplies the CPG/RBF baseline (CPGRBF + PIBB) and the episodic replay schedule used in GOLLUM's learning loop."},{"cited_title":"Nature Machine Intelligence 4(2): 169--179","cited_arxiv_id":null,"evidence_quote":"Provides the versatile modular locomotion controller that GOLLUM contrasts with, since its indirect encoding and pretraining are replaced by direct real-world learning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the ring-network CPG oscillator construction on which the sequential central pattern generator layer is based."},{"cited_title":"Neural Networks 23(4): 551--559","cited_arxiv_id":null,"evidence_quote":"Supplies the parameter-exploring policy gradient rule that GOLLUM modifies with sparse masks and gradient weighting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the episodic policy-improvement perspective underlying the dual-layer weight updates."},{"cited_title":"ACM Computing Surveys 55(13s): 1--42","cited_arxiv_id":null,"evidence_quote":"Supplies the compactness and completeness metrics used in the quantitative interpretability comparisons."},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"Defines the MORF hexapod robot used in all experiments."},{"cited_title":"Neuron 46(4): 681--692","cited_arxiv_id":null,"evidence_quote":"Motivates the use of value and observation prediction boundaries as surprise and uncertainty signals for neurogenesis."},{"cited_title":"IEEE/ASME Transactions on Mechatronics","cited_arxiv_id":null,"evidence_quote":"Provides the low-pass filter dynamics used to select the basis-layer parameters."}],"review_version":1}