{"id":"f4624f2f-85df-4963-82c0-ca1e3079ec0c","arxiv_id":"2607.03454","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Replacing adversarial motion-prior features with trajectory-optimization-derived dynamics features improves a humanoid policy's push recovery in simulation (J80 +16.7%, recovery time -47.9% vs AMP).","lead":"This paper proposes ADP, a training method that uses dynamics features — center-of-mass motion, momentum, contact forces, and contact timing — as an adversarial prior for humanoid locomotion, instead of the kinematic pose and velocity features used by AMP. In simulation, ADP recovers from pushes faster and at larger impulses than AMP while keeping velocity error lower.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ADP's recovery prior is anchored to nominal gait schedules; the ablation shows removing the prescribed contact indicator collapses success, so the claimed generality to arbitrary perturbations is not established.","rationale":"The reader identified exactly this as the weakest assumption, and the ablation on the contact indicator makes it concrete. I agree that this is the most load-bearing concern because it attacks the mechanism itself, not just the baseline. If recovery truly required leaving the nominal dynamics-feature support, the discriminator reward would suppress the needed actions, undermining the central claim. The paper's own sensitivity analysis (Table II) and ablation (Table III) are valuable and show the dynamics features are informative, but they also expose the reliance on the prescribed contact schedule. I considered the AMP baseline issue (TO-derived IK rather than mocap-trained) as a secondary concern; the paper explicitly acknowledges it in the conclusion, and it affects the strength of the head-to-head comparison but not the internal validity of the support mechanism. The proposed concrete test—augmenting the reference set with perturbed trajectories—would directly settle whether the nominal-only support is a fundamental limitation or merely a design choice. Since the reader already conditioned the verdict on such scope questions, my assessment does not change the verdict.","tokens_in":11284,"tokens_out":5738,"duration_ms":69509,"concrete_test":"Retrain ADP after augmenting D_dyn with TO-generated perturbed trajectories (e.g., SRBD TO re-solved with initial impulse disturbances to the CoM state, spanning a range of push magnitudes and phases). Compare J80 and success rate against the nominal-only ADP under the same evaluation protocol. If the augmented-reference variant yields substantially higher J80 or succeeds under mid-swing pushes where the nominal schedule is infeasible, the nominal-only support is the limiting factor and the claim of general perturbation robustness is weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The core claim is that ADP keeps policy rollouts close to the TO-derived reference support, and that this yields robust recovery from perturbations. But the reference support D_dyn (Sec. III-B) is generated by SRBD-based TO under nominal commands only, with a per-motion prescribed contact schedule {c_t^L, c_t^R}; it contains no perturbed trajectories. The discriminator reward (Eq. 10) then pulls policy windows toward this nominal support. The ablation in Table III shows that removing the binary contact indicator from the dynamics feature z_t (Eq. 4) drops success from 91.4% to 30.5%, indicating that the reward is critically dependent on enforcing the nominal contact schedule. For strong perturbations, recovery may require an extra step, altered contact timing, or a yielding posture that temporarily leaves the nominal support; the discriminator would then actively penalize the correct recovery action. The paper's evaluation sweeps impulses only up to 4.5 m/s and does not test pushes at different gait phases or magnitudes that force contact-schedule deviations. Thus the central claim that ADP encourages recovery 'even after perturbations' is only demonstrated within the narrow envelope where the nominal gait schedule remains appropriate.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes Adversarial Dynamics Priors (ADP) for humanoid locomotion control. The method replaces the kinematic features used in Adversarial Motion Priors with selected dynamics features—CoM velocity, centroidal angular momentum and its rate, normalized foot contact forces, and binary contact indicators—extracted from trajectory-optimization (TO) solutions of a single-rigid-body dynamics model. A discriminator is trained on temporal windows of these features to distinguish TO-derived reference windows from policy-generated windows, and the resulting adversarial reward is added to the task reward during PPO training. In simulation on the Unitree G1, the authors compare ADP against Vanilla RL, AMP, and a cluster-based Dynamics Reward baseline. The reported results show that ADP improves the 80%-success impulse threshold J80 by 16.7%, reduces direction-averaged recovery time by 47.9% and velocity error by 35.4% relative to AMP. Ablations show the importance of the temporal window length and of individual features, particularly the binary contact indicator. A representation-sensitivity analysis indicates that the dynamics representation responds earlier and more strongly to pushes than the joint-level kinematic representation. The paper also includes qualitative hardware demonstrations.","tokens_in":11629,"tokens_out":6676,"duration_ms":75752,"significance":"If the empirical results are robust, the contribution is significant for legged locomotion: it demonstrates that a TO-derived dynamics-feature prior can be more effective than a kinematic motion prior for perturbation recovery, and the controlled comparison using the same reference source for AMP and ADP is a strong experimental design. The representation-sensitivity analysis is a useful diagnostic. However, the central empirical claims currently lack statistical backing (no training seeds), and the perturbation envelope tested is narrow relative to the claim of general recovery. These issues must be addressed before the paper can be accepted.","major_comments":[{"comment":"The central quantitative comparisons are presented as single numbers with no error bars or number of training seeds. Because the methods are stochastic and the claimed gains (e.g., J80 115.5 vs 99.0 N·s, recovery time 2.48 vs 4.76 s) are large, the lack of variance estimates is a load-bearing gap. The IQR bands in Fig. 3 are across 32 evaluation environments from one training run and do not quantify training stochasticity. Please provide means and standard deviations over multiple seeds (e.g., 3–5), and ideally a bootstrap significance test for success rate and J80. Without this, the reader cannot judge whether the improvements are consistent.","section":"Sec. IV-A / Table I"},{"comment":"The reference dataset D_dyn is generated from TO under fixed commands and a per-motion prescribed contact schedule {c_t^L, c_t^R}. The ablation in Table III shows that removing the binary contact indicator from the dynamics feature drops direction-averaged success from 91.4% to 30.5%, demonstrating that the ADP reward is critically dependent on enforcing the nominal contact schedule. The evaluation sweeps impulses only up to Δv = 4.5 m/s and does not test pushes at different gait phases or forces that would require temporarily leaving the nominal support (extra steps, altered contact timing). Consequently, the paper's central claim that ADP 'encourages policy rollouts to remain close to the reference support, even after perturbations' is only demonstrated within an envelope where the nominal contact schedule remains a viable recovery target. Please either (i) test perturbations that forc","section":"Sec. III-B, Eq. (3); Sec. III-C, Eq. (4); Table III"},{"comment":"The default window length K=8 and the feature set are selected on the basis of the same direction-averaged success metric that is later reported as the headline result. This selection on the test metric can lead to optimistic estimates. The K-sweep shows high sensitivity (success 28.1% at K=1, 49.2% at K=4, 91.4% at K=8), so the final result hinges on this choice. Please provide evidence that the design choices generalize, e.g., by selecting K and features on a validation set of different push directions or commands, or by reporting the full ablation across a range of K values on multiple seeds. At minimum, state the selection protocol explicitly.","section":"Sec. IV-C, Fig. 6"}],"minor_comments":[{"comment":"The value of the robot mass is omitted; since J80 is defined as mΔv, please report the mass of the G1 model used in the experiments.","section":"Sec. IV-A"},{"comment":"The cluster-based Dynamics Reward baseline requires hyperparameters (τ, number of clusters, command-space bandwidth) that are not specified, making the baseline difficult to reproduce.","section":"Eq. (11) and Sec. IV"},{"comment":"The code release is promised 'on the paper website' but is not available at submission; for reproducibility, please provide code or a detailed appendix with the exact observation, architecture, and training hyperparameters.","section":"Reproducibility"},{"comment":"The normalization of 'lower-is-better metrics' is described only in the caption; please explain the normalization in the text so the reader can interpret the y-axis.","section":"Fig. 6"},{"comment":"The phrase 'direction-averaged' is sometimes hyphenated inconsistently; also check the header date '16 Jul 2026' for correctness.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the controlled comparison with AMP is a strength, but the empirical validation needs substantial strengthening: multiple training seeds, a broader perturbation envelope, and an explicit selection protocol for hyperparameters. I would like to see the revised version before making a final decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper's core idea—replacing kinematic motion priors with adversarial regularization over TO-derived dynamics features—is a real and useful variation on AMP, and the controlled comparison against an AMP baseline with the same reference source is the right way to isolate the prior representation. But the strength of the robustness claim is tied to a reference distribution built from nominal gait schedules, and the single-seed numbers in Table I leave the magnitude of the effect uncertain. It deserves a serious referee, but needs revision before the results can be taken at face value.\n\nWhat's new: prior work (Wu et al., APT-RL, DynaFlow) uses TO datasets as trajectories, decoders, or planning constraints; ADP uses a window discriminator over SRBD-derived dynamics features—CoM velocity, centroidal momentum, normalized contact forces, binary contact states—conditioned on command and tracking error. That's a distinct prior. The paper also controls the reference source cleanly: AMP and ADP use the same TO-derived motions, so the comparison isolates feature space, not data quality. The ablation is informative, even sobering: removing the binary contact indicator drops success from 91.4% to 30.5%, showing the reward is largely enforcing a prescribed contact schedule. The representation-sensitivity analysis (Table II) complements this by showing dynamics features react faster to pushes than joint kinematics.\n\nSoft spots, in order. First, the reference set contains no perturbed trajectories; the discriminator pulls the policy back to the nominal contact schedule. That is likely fine for the tested impulses (up to Δv=4.5 m/s), but the 'even after perturbations' claim is only demonstrated within the envelope where the nominal gait remains the right recovery target. A push at a different gait phase that calls for an extra step or altered contact timing could be penalized. The paper should either widen the reference set or narrow the claim. Second, Table I reports single numbers without seeds or error bars; given the 47.9% recovery-time gap, the reader can't judge whether the margin is meaningful. Third, K=8 and the feature set are selected using the same direction-averaged success metric, so the reported J80 is optimistic relative to a held-out impulse set. Fourth, the AMP baseline is a TO-derived IK variant, not a mocap-trained AMP; the authors acknowledge this, but it tempers the comparison.\n\nNone of this is fatal. The paper is coherent, the method is plausible, and the authors are honest about limitations. It deserves a serious referee, but the revision should include multiple seeds, released code/data, and a robustness test that pushes at different gait phases or includes perturbed examples in the reference set. I'd bring it to a reading group focused on legged locomotion, and I'd cite it once the code and seeds materialize.","headline":"A plausible and well-controlled demonstration that dynamics-feature adversarial priors improve push recovery in humanoids, but the headline robustness is anchored to a nominal gait schedule and the single-seed numbers need verification.","tokens_in":12146,"tokens_out":4148,"would_cite":false,"duration_ms":42371,"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":"Regularizing momentum and contact dynamics rather than joint poses makes humanoid policies substantially more push-resistant, the paper argues.","keywords":["adversarial dynamics priors","humanoid locomotion","perturbation recovery","centroidal momentum","contact forces","trajectory optimization","adversarial imitation","push recovery"],"falsifier":"Run ADP under a push whose optimal recovery requires a contact-timing change not in the nominal gait schedule (e.g., an extra step or a delayed swing phase) and observe whether the policy falls precisely when it must leave the reference contact schedule. Alternatively, retrain ADP with a reference set that includes perturbed windows and check whether J80 materially improves; if it does, the reference support is the limiting constraint.","tokens_in":11175,"feed_emoji":"🤖","tokens_out":6463,"duration_ms":61328,"temperature":0.7,"pith_summary":"The paper argues that a humanoid locomotion policy recovers from external pushes much better when the adversarial prior guiding its behavior is defined over dynamics features — center-of-mass velocity, centroidal angular momentum, normalized foot contact forces, and binary contact indicators over a short time window — instead of the joint-level kinematic features used by existing motion priors. The authors build a reference set by optimizing single-rigid-body trajectories offline, then train a discriminator to distinguish policy-generated windows from reference windows, rewarding the policy for staying close to the reference distribution after an impulse. On the simulated G1 humanoid they report that this 'adversarial dynamics prior' raises the 80%-success impulse threshold by 16.7% over the kinematic AMP baseline and cuts recovery time and velocity error roughly in half, with qualitative hardware confirmation. The significance is that push recovery hinges on momentum regulation and contact timing, quantities that kinematic features barely expose, so a dynamics-based prior gives the policy a more direct physical target.","feed_headline":"Dynamics prior makes humanoids recover from pushes twice as fast","feed_subtitle":"Adversarial regularization over momentum and contact timing beats kinematic imitation at absorbing hard pushes.","key_machinery":"The load-bearing piece is the adversarial dynamics prior over a temporal window: a discriminator that classifies eight-timestep windows of normalized dynamics features (heading-frame CoM velocity, centroidal angular momentum and its planar rate, weight-normalized foot contact forces, and binary contact indicators, conditioned on the velocity command and tracking error) as reference-like or policy-like, and a PPO policy rewarded for windows the discriminator scores as reference-like. The reference distribution is generated offline by single-rigid-body trajectory optimization, so the prior encodes physically consistent momentum–force–contact relationships rather than a particular kinematic sty","core_discovery":"The central claim is that replacing the kinematic motion-style prior in adversarial imitation with a dynamics-feature prior — CoM velocity, centroidal angular momentum, normalized contact forces, and binary contact indicators over a temporal window — yields substantially better post-push recovery without requiring reference poses, phase, or end-effector tracking. The authors construct the reference set from single-rigid-body trajectory optimization, train a discriminator on windowed dynamics features, and show that the resulting reward pulls policy rollouts back toward the reference support after a push. Quantitatively, they report a J80 of 115.5 N·s, a 91.4% direction-averaged success rate,","pith_inferences":["Because the TO reference set contains only nominal, unperturbed trajectories, the recovered behavior is defined as a pull-back to nominal dynamics. If large pushes require a non-nominal step or contact schedule, the discriminator reward may suppress the correct recovery action — a testable boundary of the method.","The binary contact indicator is effectively the prescribed gait schedule; the ablation suggests the discriminator leans heavily on gait phase. This raises the question of whether the prior is learning a dynamics distribution or a timing schedule, and whether the benefit persists under pushes that demand a phase change.","The paper compares against an AMP baseline built from the same TO-derived reference, not a high-quality mocap-based prior. A richer kinematic reference might shrink the gap, though dynamics features would likely still show earlier transient exposure.","A direct extension would be to add a small number of perturbed reference windows (e.g., from push-recovery trajectory optimization) and test whether J80 increases further; if it does, the reference support is the limiting factor."],"forward_implications":["Policies trained with ADP withstand impulses roughly 17% stronger than those trained with a kinematic motion prior before crossing an 80% success threshold.","Direction-averaged recovery time after a 3 m/s impulse push drops from 4.76 s to 2.48 s, and velocity tracking error from 1.30 to 0.84 m/s, relative to AMP.","Dynamics features expose a perturbation within 20 ms (6-fold deviation) versus 160 ms for kinematic features, so they can serve as an early detection signal for balance failures.","Contact timing and centroidal momentum are the most critical feature groups; removing the contact indicator or momentum degrades success from 91.4% to 30.5% or 45.3%, respectively.","No reference poses, phase variables, or end-effector tracking are needed, so the prior can be built purely from trajectory optimization and used for robots without motion-capture data."],"fun_headline_variants":["Dynamics prior slashes push recovery time by 48%","ADP: momentum-aware prior beats kinematic imitation on pushes","Humanoid ADP: 48% faster recovery after hard pushes","Dynamics-based prior boosts push impulse threshold by 17%","ADP: dynamics features, not poses, stabilize humanoids under pushes"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The TO reference set is built only for nominal locomotion modes under fixed commands and contains no perturbed trajectories; the recovery mechanism assumes the desired post-push behavior is to be pulled back into the nominal dynamics-feature support, so if strong perturbations require temporarily leaving that support (e.g., an extra step or a yielding posture), the discriminator reward could suppress the very actions needed to recover.","fun_headline_variants_meta":{"raw":{"variants":["Dynamics prior slashes push recovery time by 48%","ADP: momentum-aware prior beats kinematic imitation on pushes","Humanoid ADP: 48% faster recovery after hard pushes","Dynamics-based prior boosts push impulse threshold by 17%","ADP: dynamics features, not poses, stabilize humanoids under pushes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000804,"raw_usage":{"total_tokens":3353,"prompt_tokens":712,"completion_tokens":2641,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":2554}},"tokens_in":456,"tokens_out":2641,"duration_ms":19640,"temperature":1.0,"reasoning_tokens":2554,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T08:50:04.636722+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ADP under a push whose optimal recovery requires a contact-timing change not in the nominal gait schedule (e.g., an extra step or a delayed swing phase) and observe whether the policy falls precisely when it must leave the reference contact schedule. Alternatively, retrain ADP with a reference set that includes perturbed windows and check whether J80 materially improves; if it does, the reference support is the limiting constraint.","supporting_citations":[],"review_version":2}