{"id":"dfb49614-4497-4328-a559-7fa2370d998f","arxiv_id":"2607.20110","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"A two-stage continual-learning framework lets a generalist humanoid tracking policy acquire highly dynamic acrobatic skills while preserving its general-purpose motion capabilities.","lead":"This paper shows a humanoid robot can learn rare acrobatic skills—backflips, air twists, kip-ups—on top of a generalist motion-tracking policy without forgetting how to walk and move normally. It is a step toward humanoid controllers that combine broad competence with expert-level dynamic agility.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"AMASS Challenging test set may overlap Stage-II acquisition data; the quantitative 'unseen' dynamic-motion claim is not yet supported.","rationale":"The reader's weakest assumption (motion stratification from five rollouts) is reasonable but secondary: even a noisy split would still allocate most hard motions to D_c and easy motions to D_m, and the PACE/STAR mechanism may tolerate moderate misclassification. The paper's ablations (Mixed Training, w/o L_con, fixed lambda_con) also address related design choices. The train/test overlap issue is more load-bearing because it directly threatens the central claim's 'unseen' qualifier. The paper's quantitative evidence for unseen dynamic motions is Table VI(b)'s AMASS Challenging result; everything else for unseen dynamic motions is hardware (4 motions × 5 trials) or XtremeMotion, which is a separate dataset with higher-quality preprocessing. If AMASS Challenging overlaps D_c, the reported 90.91% is not a generalization result. The same concern does not apply to the generalist In-source/Unseen subsets: In-source is explicitly from training sources, and Unseen is a separate reconstructed video set. But for the specialist claim, a held-out split is necessary. This is an addressable empirical issue, not a logical contradiction, so the reader's CONDITIONAL verdict stands. I mark agreement as partial because the reader identified overlap as a concern but selected stratification as the weakest assumption; my stress-test prioritizes the overlap.","tokens_in":21462,"tokens_out":4883,"duration_ms":55228,"concrete_test":"Ask the authors to publish the AMASS sequence identifiers and temporal intervals for every clip in D_c and in the AMASS Challenging test set. Then recompute Table VI(b) under a strict split: remove from D_c any clip whose source sequence overlaps any AMASS Challenging test clip — including partial temporal overlap — retrain Stage II, and re-evaluate. If the held-out success remains close to 90.9%, the concern is resolved. If it drops toward the Stage-I/Fine-Tuning range (18–55%), the generalization part of the central claim is unsupported, and the paper should report only training-set specialist success.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-C defines the full set D as LAFAN1 + AMASS + in-house Xsens and forms D_c (the challenging acquisition set) from all clips that fall below the 80% completion threshold. Section VI.A.3 evaluates 'AMASS Challenging Motions' as 'difficult motions selected from AMASS.' No sentence anywhere states that these test clips, or even their source AMASS sequences, are excluded from D_c. Since D_c is built from AMASS, the default reading is that the test set overlaps the acquisition data. If the same clips (or overlapping temporal windows of the same AMASS sequences) are used in Stage-II training and in the 90.91% success result in Table VI(b), then that number is a training-set fitting result, not evidence of generalization to unseen highly dynamic motions. The abstract's 'directly executes diverse unseen highly dynamic motions' then rests on XtremeMotion (a separate dataset) and the hardware demos, which are not quantitatively controlled for novelty. This does not invalidate the relative comparison against Fine-Tuning and OmniXtreme on the same test set, but it removes the main quantitative support for the 'generalist expands to unseen dynamic skills' component of the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Extreme-RGMT, a two-stage continual learning framework for Unitree G1 whole-body motion tracking. Stage I trains a generalist base policy from LAFAN1, AMASS, and in-house Xsens data using an architecture with separated state/action/reference encoders and FSQ command quantization. Stage II freezes the base policy as a reference, splits motions into mastered and challenging sets, and trains an augmented policy with PACE, an asymmetric acquisition/consolidation objective with progress-adaptive regularization, and STAR, a fragment-level advantage resampling mechanism. The paper reports state-of-the-art generalist tracking success, high success on XtremeMotion and AMASS Challenging motions, and hardware demos for fixed-reference replay and Xsens teleoperation.","tokens_in":21745,"tokens_out":3275,"duration_ms":38544,"significance":"If the results hold, the paper makes a meaningful advance: it shows a single humanoid policy can retain broad generalist motion-tracking ability while substantially improving execution of highly dynamic, failure-prone motions, and it demonstrates this on hardware with online inertial motion-capture input. The work is strengthened by controlled retraining of ExBody2, BeyondMimic, and RGMT on the same Stage-I corpus, the use of a public external checkpoint (SONIC, OmniXtreme) as a comparison, component ablations in simulation and hardware, and five-seed error-metric reporting. The main risk is that the central 'unseen highly dynamic motions' claim is not yet supported by the quantitative specialist evaluation, because the AMASS Challenging test set appears to overlap the Stage-II acquisition set.","major_comments":[{"comment":"The AMASS Challenging Motions test set is not shown to be disjoint from the Stage-II acquisition set D_c. Section IV-C constructs D_c from all AMASS clips whose completion rate under the base policy is below 80%, while Section VI.A.3 evaluates 'difficult motions selected from AMASS.' No sentence states that these test clips, or the AMASS sequences they are drawn from, are excluded from D_c. If the same clips or temporally overlapping windows are used in Stage-II training, the 90.91% specialist success on AMASS Challenging Motions is training-set fitting rather than evidence of generalization to unseen dynamic motions. The authors must report the exact overlap, or re-evaluate on held-out AMASS sequences. The same concern applies to hardware 'AMASS Replay' and the Xsens teleop motions, which may also be drawn from D_c. This is load-bearing for the abstract and conclusion claims about execu","section":"Sec. IV-C and Sec. VI.A.3, Table VI(b)"},{"comment":"The motion stratification that defines D_m and D_c is based on only five stochastic rollouts per clip and an 80% completion-rate threshold. With five binary rollouts, the completion rate is quantized to 0, 20, 40, 60, 80, or 100 percent, so the 80% boundary is coarse. The entire Stage-II curriculum—which motions are consolidated versus acquired—inherits the noise and metric choice of this estimate. The paper provides no sensitivity analysis with respect to rollout count, threshold value, or seed. I request an analysis showing the D_c/D_m assignment is stable, or at least that the final simulation and hardware results are not sensitive to reasonable perturbations of the stratification procedure.","section":"Sec. IV-C"},{"comment":"Success rates are the headline metric for both the generalist and specialist claims, but they are reported as point values without variance, despite the paper stating that all results are averaged over five random seeds. The error metrics include standard deviations; the success rates should also report per-seed values, confidence intervals, or a statistical comparison. This is especially important for differences such as Fine-Tuning (54.55%) versus Extreme-RGMT Full (90.91%) on AMASS Challenging Motions, and Extreme-RGMT Stage I versus Full on the Unseen Motion set. Currently the reader cannot judge whether the reported success-rate differences are significant at the five-seed scale.","section":"Table VI and Sec. VI.A.1"}],"minor_comments":[{"comment":"The phrase 'directly executes diverse unseen highly dynamic motions' overstates the evidence if the AMASS Replay and Xsens hardware motions are drawn from the training corpus or from D_c. Please rephrase to 'motions not explicitly held out' or specify which evaluations are truly unseen.","section":"Abstract and Sec. VII"},{"comment":"The name 'AMASS Challenging Motions' is easily confused with the training set D_c. Consider renaming the evaluation subset, e.g., 'AMASS Held-Out Challenging Motions' if it is in fact held out, and explicitly state its relationship to D_c.","section":"Sec. VI.A.3"},{"comment":"The time-step axes in Fig. 8 have no labels or units, and the two panels are not annotated. Please add axis labels and clarify whether the plot shows a specific seed, a representative rollout, or an aggregated statistic.","section":"Fig. 8"},{"comment":"Table VII reports 'In-house Xsens Motions' while Table VIII reports 'Xsens Teleop.' Please state explicitly that Table VII is simulation and Table VIII is hardware, to avoid ambiguity.","section":"Table VII vs Table VIII"},{"comment":"Please clarify whether the five randomized rollouts used for stratification follow the same randomization seed protocol as the evaluation rollouts, and whether the 80% threshold is applied to the mean or to each seed independently.","section":"Sec. IV-C"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically strong and the proposed mechanisms are plausible, but the headline 'unseen highly dynamic motions' claim is not yet substantiated because of the suspected training/test overlap in the AMASS Challenging evaluation. This is fixable by disclosing the overlap or re-running on truly held-out AMASS sequences, and by reporting success-rate variance. I recommend major revision rather than rejection: the framework and ablations are valuable, and the central contribution can likely be made sound with focused additional analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the engineering is real: the two-stage setup with PACE and STAR produces a humanoid policy that tracks a broad vocabulary and also does backflips, kip-ups, and aerial twists, and on hardware it follows online Xsens input. The ablations are the right ones, and the external checkpoints (SONIC, OmniXtreme) give the comparisons teeth. Second, there is a likely train/test overlap that undermines the headline \"unseen\" claim. The challenging acquisition set D_c is defined in Section IV-C as every clip from the full D (LAFAN1 + AMASS + Xsens) that fails the 80% completion bar. The AMASS Challenging test set in Section VI.A.3 is \"difficult motions selected from AMASS.\" Nothing says those test clips were held out of D_c. So the 90.91% success in Table VI(b) may be a fitting result. That does not sink the paper—XtremeMotion is a separate dataset and the hardware demos are genuinely unseen—but it does mean the quantitative generalization claim on AMASS is not supported as written.\n\nWhat's genuinely new: the stratification step itself, the role-specific consolidation with a frozen reference policy, and the advantage-weighted fragment resampling. The ablations show each piece earns its keep, and STAR's gain is largest on noisy Xsens data, which fits the argument. The \"first online inertial teleoperation of acrobatic motions\" claim is plausible from the videos and the hardware table.\n\nSoft spots, in order. (1) The overlap above; the authors need to either exclude AMASS Challenging clips from D_c or re-run with a properly held-out set. (2) Success rates are reported as point estimates only. Five seeds should be enough to give std or CIs; without them I can't tell whether the 90.9 versus 89.6 differences are real. (3) No code or data. For a method with this many knobs (consolidation weight, gain, thresholds, resampling ratios), that is a limitation. (4) The stratification uses five rollouts and an 80% threshold; the paper doesn't test sensitivity to that split. I consider that minor relative to the others. The limitations section is honest about drift and narrow generalization, which I read as a good sign.\n\nBottom line: this deserves a serious referee. The system works, and the idea of progressive, role-separated acquisition is worth engaging. But the test-set overlap and the missing variance need to be addressed before the core claim is fully convincing. I'd send it out and ask for a revision, not a desk rejection.","headline":"Strong empirical system with a real data-leakage question: the AMASS Challenging test may be in the Stage-II training set, so the 'unseen' claim needs a fix before I'd trust the headline.","tokens_in":115,"tokens_out":2562,"would_cite":true,"duration_ms":74135,"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":"A two-stage continual-learning framework lets a single humanoid controller track both everyday motions and acrobatic high-dynamics moves, even from noisy inertial motion-capture input.","keywords":["humanoid whole-body control","motion tracking","continual learning","reinforcement learning","highly dynamic skills","teleoperation","inertial motion capture"],"falsifier":"Retrain Stage II using the same base policy but with motion stratification computed from 50 rollouts per clip instead of 5, or with thresholds of 70% and 90%: if specialist success on the challenging set drops substantially or generalist retention degrades, the curriculum is riding on the luck of a five-rollout estimate. On hardware, run the teleoperation policy with operators whose motion timing deviates more than roughly 20% from the training clips and see whether the completion rate collapses.","tokens_in":21325,"feed_emoji":"🤸","tokens_out":2907,"duration_ms":28605,"temperature":0.7,"pith_summary":"This paper claims that the long-standing trade-off in humanoid control between broad generalist tracking and reliable execution of rare, highly dynamic motions can be overcome by structuring training as a two-stage progression: first a generalist base policy over diverse motion data, then an asymmetric second stage where challenging motions drive skill acquisition and mastered motions are consolidated to prevent forgetting. The key mechanisms are PACE, which splits training into acquisition and consolidation roles with a progress-adaptive constraint, and STAR, which prioritizes scarce high-advantage trajectory fragments within difficult temporal segments. If correct, a single policy can retain roughly 99% success on seen general motions and 97% on unseen motions while raising highly dynamic motion success from about 18-21% to 91-100% in simulation, and executing acrobatics online from inertial mocap on hardware. This matters because it moves humanoid control from offline replay of cleaned reference trajectories toward real-time human-driven teleoperation of athletic skills.","feed_headline":"One policy handles backflips and everyday humanoid motion","feed_subtitle":"Two-stage continual learning keeps ~99% generalist tracking while adding acrobatic skills at ~91% success","key_machinery":"The paper's central machinery is PACE (Progressive Acquisition and Consolidation for Expansion), which asymmetrically allocates parallel environments between acquisition on a challenging motion set and consolidation on a mastered set, coupling a PPO acquisition loss with a reference-policy action-alignment loss ∥aθ(s) − aref(s)∥² whose weight λcon rises adaptively as the fraction of valid acquisition samples increases. STAR (Segment-Aware Trajectory Advantage Resampling) converts bin-level difficulty priors into transition-level difficulty weights, normalizes advantages separately for high-difficulty versus remaining transitions, ranks contiguous trajectory fragments by average raw advantage","core_discovery":"The central claim is that highly dynamic skill acquisition and generalist capability retention need not be mutually exclusive if the training curriculum is organized asymmetrically by role. A consolidation branch constrains policy drift on motions the base policy already masters, using reference-policy regularization whose strength grows with training progress, while an acquisition branch focuses exclusively on challenging motions using difficulty-aware sampling and advantage-prioritized trajectory-fragment resampling. On the paper's own terms, the resulting augmented policy is the first generalist humanoid controller that directly executes diverse unseen highly dynamic motions under online","pith_inferences":["The same two-role curriculum idea could transfer to other contact-rich and dynamic robot tasks—manipulation, loco-manipulation, or parkour—where decisive learning signals are concentrated in short segments and failure samples dominate the buffer.","The motion-stratification step (five rollouts, 80% threshold, 0.2 m root-height failure) is the main sensitivity: the entire Stage-II curriculum inherits the noise of that label, so a confidence-weighted split or a robustness study over rollout counts and thresholds would clarify how reproducible the reported gains are.","The gains on unseen motions hint that the framework may generalize beyond the training distribution, but the paper's own limitation statement concedes that motions with unfamiliar coordination patterns or timing still fail, so the practical boundary lies where the training data have no near-neighbor coverage."],"forward_implications":["A single policy can span everyday motions and acrobatic skills, removing the need to switch between a generalist and a specialist controller depending on the task.","Online teleoperation of highly dynamic motions becomes feasible from low-quality inertial motion-capture references, extending humanoid control from offline replay to real-time human-driven execution.","Training on scarce, failure-prone data can be made more sample-efficient by concentrating optimization on high-advantage trajectory fragments in difficult temporal segments, not just on difficult bins.","The consolidation mechanism offers a reusable template for continual robot learning in which new skills are built on stable base competencies without catastrophic forgetting."],"fun_headline_variants":["One policy does backflips and everyday walking","Humanoid learns flips, keeps daily skills","Asymmetric training: generalist plus extreme moves","Continual learning for robust humanoid acrobatics","From basics to backflips: one humanoid controller"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The split of motions into 'mastered' and 'challenging' rests on the completion rate from just five stochastic rollouts under the base policy with an 80% threshold, and the entire second-stage curriculum is built on that noisy, metric-sensitive label.","fun_headline_variants_meta":{"raw":{"variants":["One policy does backflips and everyday walking","Humanoid learns flips, keeps daily skills","Asymmetric training: generalist plus extreme moves","Continual learning for robust humanoid acrobatics","From basics to backflips: one humanoid controller"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000205,"raw_usage":{"total_tokens":1213,"prompt_tokens":709,"completion_tokens":504,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":430}},"tokens_in":453,"tokens_out":504,"duration_ms":4944,"temperature":1.0,"reasoning_tokens":430,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T10:44:09.551751+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain Stage II using the same base policy but with motion stratification computed from 50 rollouts per clip instead of 5, or with thresholds of 70% and 90%: if specialist success on the challenging set drops substantially or generalist retention degrades, the curriculum is riding on the luck of a five-rollout estimate. On hardware, run the teleoperation policy with operators whose motion timing deviates more than roughly 20% from the training clips and see whether the completion rate collapses.","supporting_citations":[],"review_version":1}