{"id":"f2734d4a-fdb1-4e96-a910-4104268c0926","arxiv_id":"2508.19926","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"FARM combines frame-accelerated augmentation with a residual mixture-of-experts to track high-dynamic humanoid motions, cutting tracking failures by 42.8% on a new HDHM benchmark.","lead":"This paper adds a residual mixture-of-experts and accelerated-frame training data to a frozen humanoid controller, improving tracking of explosive motions like cartwheels and martial arts. The authors also introduce HDHM, a 3,593-clip benchmark for high-dynamic humanoid control, and report a 42.8% drop in tracking failures versus the baseline.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"HDHM benchmark appears constructed from FC failure cases, so the 42.8% failure-rate improvement may be inflated by selection bias; Appendix A contradicts Table 1.","rationale":"The reader's weakest_assumption identifies exactly this benchmark-construction bias, and I concur. The paper's central claim is the 42.8% failure-rate reduction on HDHM. If HDHM is curated by selecting clips where FC fails, then FC's performance is artificially low and FARM's advantage is not a fair measure of generalization. The internal inconsistency with Table 1 (FC 92.3% success on HDHM) strengthens the concern: either the curation description is misleading or the reported success rate is incompatible with the described procedure. This is a concrete, load-bearing flaw that can be checked by re-running the selection pipeline and/or by evaluating on a randomly sampled high-dynamic set. The reader's CONDITIONAL verdict remains appropriate: the paper should not be accepted without addressing this issue, but the method and ablations are potentially valuable, so rejection is too strong. I agree with the reader's assessment, so no verdict change is needed.","tokens_in":11607,"tokens_out":3528,"duration_ms":38950,"concrete_test":"Reconstruct the HDHM selection procedure: run the released FC weights on the full candidate pools (AIST++, EMDB, Motion-X Kungfu, Text-Convert, Video-Convert) and determine which final HDHM clips were included because FC produced a tracking failure. If >80% of HDHM clips were selected due to FC failure, the success-rate comparison against FC is biased. Also recompute FC's success rate on the final HDHM list using the paper's stated success criterion; if it differs substantially from the reported 92.3%, the internal inconsistency is confirmed. Independently, evaluate FARM and FC on a random (non-FC-filtered) subset of high-dynamic motions; if the 42.8% reduction shrinks or reverses, selection bias is the cause.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—FARM reduces tracking failure rate by 42.8% on HDHM—rests on HDHM being a fair, unbiased high-dynamic benchmark. Appendix A states: 'we first run the FC baseline policy on all candidate sequences and collect those with significant tracking failures.' This makes the evaluation set dependent on the baseline being compared: HDHM is enriched for FC failures. If the dataset literally consists of clips where FC fails, then FC's success on HDHM should be far below the reported 92.3% in Table 1. The only way both statements can be true is if 'significant tracking failures' means something weaker than the clip-level success metric, or the curation was not applied as described. Either way, the comparison is not an unbiased estimate of generalization. Moreover, FARM is trained on failure cases mined from AMASS using FC, so its training distribution is aligned with an FC-failure-selected evaluation distribution, which can inflate the apparent improvement. The absence of error bars further weakens the reported effect. This selection bias is load-bearing because the headline result is the primary evidence for FARM's contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FARM, a physics-based humanoid controller that combines frame-accelerated augmentation, a frozen/partially frozen FC base controller, and a residual mixture-of-experts (MoE) module with a speed-aware router and dynamic expert assignment. Training uses hard samples mined from AMASS by running the FC baseline at 1.25x speed, then applying random 1.0–1.5x acceleration. The authors curate a new HDHM benchmark of 3,593 high-dynamic clips from five sources and report that FARM reduces the tracking failure rate on HDHM by 42.8% (from 7.7% to 4.4%) and lowers MPJPEg by 14.6% (from 111.3 mm to 95.0 mm) relative to FC, while preserving near-perfect accuracy on AMASS. Ablations examine the contribution of each component, and a failure-data fine-tuning strategy is shown to reduce training time.","tokens_in":11968,"tokens_out":5355,"duration_ms":59644,"significance":"If the claimed results are valid, the paper makes a useful contribution to physics-based humanoid control: it identifies a concrete limitation of existing AMASS-trained controllers on high-dynamic motions and proposes a modular architectural fix (residual MoE with speed-aware routing) plus a data-augmentation scheme that requires only low-dynamic motion data. The curated HDHM benchmark, if released in an unbiased form, would also be a valuable community resource. The authors report four-seed averages and provide a detailed ablation table, which is a strength. However, the central quantitative claims depend on the integrity of the HDHM evaluation set, and that set is described in Appendix A as being selected using the very baseline being compared. This, together with an inconsistency in the DEA formulation, prevents the current manuscript from supporting its headline conclusions.","major_comments":[{"comment":"HDHM curation is not independent of the baseline. Appendix A says 'we first run the FC baseline policy on all candidate sequences and collect those with significant tracking failures.' If this uses the same failure criterion as §3.1 (MPJPEg > 0.5 m at any frame), then FC should fail on essentially all HDHM clips, contradicting the 92.3% success rate in Table 1. If 'significant tracking failures' means a weaker/different criterion, it must be specified precisely. As written, the evaluation set is enriched for FC failures, and FARM is additionally trained on FC-mined failure cases (§3.1, Table 4), so the reported 42.8% failure-rate reduction is not an unbiased estimate. Please re-curate HDHM without using FC as a filter, report results on the full candidate pool before filtering, or provide another unbiased benchmark; also report per-seed results/error bars.","section":"Appendix A; Table 1; §3.1"},{"comment":"The DEA formulation as written cannot activate more than E-1 experts. With p∈R^E and k=argmax(p), k ranges over {0,...,E-1}, so the maximum number of active experts is E-1. Since §B.2 reports E=2 experts, the model should only ever activate 0 or 1 experts. This contradicts Figure 6, which shows frames with 2 active experts, and Figure 7, which states that over half of HDHM frames involve one or two experts. Either the router outputs E+1 logits (including a no-expert option) or the activation rule is described incorrectly. This is load-bearing for the DEA 'adaptive capacity' claim, and the notation must be corrected and matched to the implementation.","section":"§3.3, Eq. (4); §B.2"},{"comment":"The results are described as averaged over four random seeds, but no standard deviations, confidence intervals, or per-seed values are given. For many of the reported differences (e.g., EMDB success 95.5% vs 97.7%, Text-Convert MPJPEg 76.9 vs 71.2 mm), one cannot assess whether the differences are significant. Quantitative claims such as 'consistently outperforms baselines across all subsets' need variance information, especially because the benchmark construction already introduces a baseline-dependent selection step.","section":"Table 1; §4.2"}],"minor_comments":[{"comment":"Typo: 'Issac Lab' should be 'Isaac Lab'.","section":"§3.1"},{"comment":"The caption says 'training parameters are shown in Figure 5', but the hyperparameters are in Table 5; the cross-reference is incorrect.","section":"Appendix B.3, Table 5"},{"comment":"The parenthetical percentage changes are ambiguous. For example, 92.3% to 95.1% is shown as '+3.0%', which is closer to a relative change (3.0%) than an absolute percentage-point change (+2.8 pp). Please define the convention used.","section":"Table 3"},{"comment":"The paper promises code and dataset release, but the Appendix A description of the filtering procedure is not fully reproducible: the 'significant tracking failures' criterion is not quantified, and the manual filtering thresholds for 'mild imperfections' are subjective. Please provide the exact filtering code/checklist or a more operational definition.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The benchmark circularity is the central risk. If the authors cannot re-curate or report on an FC-independent evaluation set, the paper's primary comparison will remain unreliable regardless of the architectural contribution. The DEA notation issue also needs to be fixed before the method can be assessed accurately."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Briefly: the paper has a sensible core idea, but the evaluation benchmark appears to have been built from the baseline's failures, which would make the headline numbers misleading. That needs to be sorted out before trusting the results.\n\nWhat's new: the frame-acceleration augmentation is a clean, simple way to generate high-dynamic training data from low-dynamic motion capture. The residual MoE with speed-aware routing and dynamic expert assignment is a reasonable way to add capacity without wrecking the pretrained base. The ablations in Table 3 are genuinely informative; they show that the residual design matters and that full fine-tuning hurts low-dynamic performance. The curated HDHM dataset, with 3593 clips from five sources, could be a useful community resource if its construction is corrected and documented.\n\nThe soft spots are real. Appendix A says HDHM was created by running the FC baseline on all candidate sequences and keeping those with 'significant tracking failures,' then manually filtering for physical plausibility. If that is what happened, FC should fail on nearly all of HDHM, not on 7.7% as reported in Table 1. Either the description is wrong, or 'significant tracking failures' means something much weaker than clip-level failure. In either case, the evaluation set is not independent of the baseline, so the 42.8% failure-rate reduction and 14.6% MPJPEg improvement are not trustworthy as reported. The training procedure makes it worse: FARM is trained on failure cases mined from AMASS with the same FC controller, so the training distribution is aligned with the evaluation distribution in a way that can inflate the apparent gain. I don't see error bars anywhere, and code/data are not yet available, so this cannot be checked independently right now.\n\nThe reader's report says 'conditional' and I'd agree. The method itself is plausible and the ablations are solid enough to warrant a closer look. The right response is not to reject it out of hand, but to send it to review with the dataset construction question front and center. If the authors can clarify or fix the benchmark, this becomes a decent contribution to the humanoid control subfield. If not, the headline result should not be taken at face value.","headline":"A sensible residual MoE method for high-dynamic humanoid control, but the benchmark appears to be built from the baseline's failures—so the headline numbers need a careful second look.","tokens_in":12372,"tokens_out":3268,"would_cite":false,"duration_ms":35808,"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 residual mixture-of-experts with frame-accelerated augmentation lets one physics-based humanoid controller track explosive, high-dynamic motions without losing its accuracy on ordinary motions.","keywords":["physics-based humanoid control","motion imitation","residual mixture-of-experts","frame-accelerated augmentation","reinforcement learning","high-dynamic motions","HDHM benchmark"],"falsifier":"Run FARM and the FC baseline on an independently curated set of high-dynamic motions that no controller was used to screen; if FARM does not reproduce roughly the 42.8% failure reduction and 14.6% MPJPEg improvement there, the benchmark's baseline-screened selection is the source of the reported gains.","tokens_in":11588,"feed_emoji":"🏃","tokens_out":5064,"duration_ms":58610,"temperature":0.7,"pith_summary":"FARM aims to show that a single physics-based humanoid controller can handle both everyday motions and explosive, high-dynamic actions, without per-motion retraining. The paper's core idea is to synthesize high-dynamic training data by resampling existing low-dynamic mocap at faster frame rates, then attach a residual mixture-of-experts module to a frozen base controller so extra network capacity appears only when motion intensity demands it. On a newly curated high-dynamic benchmark, the paper reports a 42.8% reduction in tracking failures and a 14.6% lower global per-joint position error compared with the strong FC baseline, while preserving 100% success on low-dynamic AMASS motions. If correct, this would make unified humanoid controllers substantially more useful for robotics and animation, where fast transitions and loss of balance are common failure points.","feed_headline":"Faster-frame training cuts high-dynamic humanoid failures by 42.8%","feed_subtitle":"A residual mixture-of-experts controller tracks explosive motions while keeping near-perfect accuracy on everyday motions.","key_machinery":"The load-bearing mechanism is the residual mixture-of-experts: a frozen pretrained controller is augmented by expert transformer adapters whose outputs enter through zero-initialized projections, so the residual path starts as a no-op. The speed-aware router assigns samples to one of three speed bands via a cross-entropy auxiliary loss, and dynamic expert-assignment activates the top k experts where k is the router's argmax, with weights formed by cumulative router probabilities. Frame-accelerated augmentation supplies the high-dynamic signal: it resamples each training clip at a random 1.0–1.5× speed and feeds the accelerated poses to the controller at the original 30 Hz rate, widening inte","core_discovery":"The central claim is that high-dynamic humanoid motion tracking can be improved without degrading low-dynamic performance by combining frame-accelerated augmentation with a residual mixture-of-experts architecture. The base FC controller is kept frozen, and a residual MoE adds expert corrections only for high-velocity segments. A speed-aware router sorts training samples into speed bands and supervises the router with those labels, while dynamic expert-assignment activates zero, one, or more experts depending on the router's confidence, so easy motions run through the base controller alone and hard motions recruit more capacity. The paper also introduces HDHM, a 3,593-clip benchmark of physi","pith_inferences":["Left implicit by the paper: frame-accelerated augmentation is a generic stress test for any tracking policy, so one testable extension is mapping a controller's failure boundary across continuous speed factors rather than only the 1.0–1.5× range used here.","The speed-aware routing idea could extend beyond joint speed to other continuous difficulty axes, such as terrain roughness, external perturbation magnitude, or task complexity, whenever samples can be labeled along that axis.","Because HDHM clips were selected by running the FC baseline and keeping its failures, an independent high-dynamic benchmark not screened by any baseline would be needed to separate FARM's genuine generalization from selection bias.","For real-robot deployment, the residual design might serve as an online adaptation layer on top of a deployed base controller, but the paper does not test sim-to-real transfer or latency, so that remains an open extension."],"forward_implications":["High-dynamic control skills can be mined from predominantly low-dynamic data via speed resampling, reducing the need for new explosive-motion capture collections.","Freezing the base controller while adding a residual MoE protects existing low-dynamic accuracy; the reported AMASS success stays at 100% with less than 2 mm added MPJPEg.","Adaptive expert activation means compute follows motion intensity: roughly 80% of low-dynamic frames use only the base controller, while over half of HDHM frames engage one or two experts.","Failure-case fine-tuning concentrates training on informative hard samples, cutting training time from 35 hours to 6 hours while slightly improving both HDHM and AMASS-test metrics.","The HDHM benchmark provides the first public high-dynamic evaluation set, covering dance, sports, kungfu, text-generated, and video-converted motions, allowing standardized comparison for future controllers."],"supporting_citations":[{"why":"Defines the universal humanoid tracking problem, the success metric, and the failure threshold (global MPJPE exceeding 0.5 m) used for hard-sample mining.","marker":"Luo et al. 2021"},{"why":"Supplies the FC controller used as the base model and main baseline, including its network architecture, reward terms, and training setting.","marker":"Tessler et al. 2024"},{"why":"Introduces PHC, a prior progressive multiplicative controller that FARM is compared against and builds upon.","marker":"Luo et al. 2023"},{"why":"Provides PHC+, a state-of-the-art baseline, and the AMASS filtering/splitting procedure that the paper follows.","marker":"Luo et al. 2024"},{"why":"Supplies the AMASS motion-capture corpus, the low-dynamic training data from which failure cases are mined.","marker":"Mahmood et al. 2019"},{"why":"Motivates the zero-initialized residual projection design that keeps the MoE path from disturbing the pretrained controller's initial behavior.","marker":"Zhang, Rao, and Agrawala 2023"},{"why":"Provides the Isaac Lab simulation environment used for reinforcement learning training and evaluation.","marker":"Mittal et al. 2023"}],"fun_headline_variants":["FARM cuts high-dynamic humanoid failures 42.8%, errors 14.6%","Frame acceleration + residual MoE: 42.8% fewer humanoid failures","New HDHM benchmark: FARM reduces failures by 42.8%","42.8% fewer failures: FARM's frame-accelerated MoE for humanoid control"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The HDHM benchmark clips are collected by running the FC baseline and keeping motions it fails on, so the claimed improvements may reflect fitting to that specific selection rather than a general advantage on all high-dynamic motions.","fun_headline_variants_meta":{"raw":{"variants":["FARM cuts high-dynamic humanoid failures 42.8%, errors 14.6%","Frame acceleration + residual MoE: 42.8% fewer humanoid failures","New HDHM benchmark: FARM reduces failures by 42.8%","42.8% fewer failures: FARM's frame-accelerated MoE for humanoid control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00076,"raw_usage":{"total_tokens":3228,"prompt_tokens":777,"completion_tokens":2451,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":2357}},"tokens_in":521,"tokens_out":2451,"duration_ms":17234,"temperature":1.0,"reasoning_tokens":2357,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:20:45.835220+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FARM and the FC baseline on an independently curated set of high-dynamic motions that no controller was used to screen; if FARM does not reproduce roughly the 42.8% failure reduction and 14.6% MPJPEg improvement there, the benchmark's baseline-screened selection is the source of the reported gains.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the universal humanoid tracking problem, the success metric, and the failure threshold (global MPJPE exceeding 0.5 m) used for hard-sample mining."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the FC controller used as the base model and main baseline, including its network architecture, reward terms, and training setting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces PHC, a prior progressive multiplicative controller that FARM is compared against and builds upon."},{"cited_title":"M.; and Xu, W","cited_arxiv_id":null,"evidence_quote":"Provides PHC+, a state-of-the-art baseline, and the AMASS filtering/splitting procedure that the paper follows."},{"cited_title":"F.; Pons-Moll, G.; and Black, M","cited_arxiv_id":null,"evidence_quote":"Supplies the AMASS motion-capture corpus, the low-dynamic training data from which failure cases are mined."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the zero-initialized residual projection design that keeps the MoE path from disturbing the pretrained controller's initial behavior."},{"cited_title":"L.; Singh, R.; Guo, Y.; Mazhar, H.; Mandlekar, A.; Babich, B.; State, G.; Hutter, M.; and Garg, A","cited_arxiv_id":null,"evidence_quote":"Provides the Isaac Lab simulation environment used for reinforcement learning training and evaluation."}],"review_version":1}