{"id":"328c4484-fd01-4d1f-9396-ba97335b5cb2","arxiv_id":"2608.02385","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A proprioception-gated mixture-of-experts policy lets a humanoid robot both track continuous motion and recover from falls with human-like get-up behavior, without any runtime recovery reference or command.","lead":"StableMimic trains a single humanoid robot control policy that both tracks a motion command and autonomously recovers from falls by blending two specialized sub-policies. It outperforms existing trackers on LAFAN1 dance tracking and recovers from 100/100 scripted pushes in simulation, with qualitative hardware demos on a Unitree G1.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Gate trained on privileged rollout labels may not transfer to push-to-fall states outside the sampled support Rδ; 100/100 on one protocol is insufficient evidence.","rationale":"The reader's verdict is CONDITIONAL, and the identified weakest assumption is the transfer of the proprioceptive gate from privileged rollout labels to deployment. My stress-test sharpens this concern: the training support Rδ is defined by static perturbed get-up states with tight bounds, while the deployment push protocol produces dynamic fall states with much larger velocities and contact configurations. The paper's own scope statement in §III-B excludes states outside Rδ, yet the recovery experiments likely operate outside it. Thus the 100/100 recovery claim is not grounded in the training distribution; it is an untested generalization. No evidence of variance is provided, so the single deterministic run cannot rule out chance. This is the most load-bearing point because if the gate misclassifies at deployment, the entire single-policy architecture collapses—either the tracking expert chases an infeasible command or the recovery expert activates during normal tracking. I agree with the reader that the paper is otherwise well-scoped and the method is plausible. The proposed concrete test would directly probe gate confidence and success on out-of-support states, settling whether the transfer premise holds. I therefore keep the verdict as CONDITIONAL, unchanged.","tokens_in":10469,"tokens_out":6923,"duration_ms":73211,"concrete_test":"Generate 200 new push-to-fall trials with forces sampled from 400–800 N, 8 evenly spaced horizontal directions, and pushes occurring at randomly chosen dance phases. For each state visited, record (a) the gate's softmax routing weights and (b) the minimum normalized distance to the stored get-up reference library using the Table V perturbation metrics. If the routing confidence (max weight) falls below 0.9 at any out-of-support state, or if recovery success drops below 95%, the transfer premise fails. Report the fraction of rollout states that fall outside Rδ to determine whether the experiments are inside the paper's stated scope.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—one proprioception-gated policy handles both tracking and post-fall recovery without a runtime reference—rests on the gate learned in §III-C(a) from rollout-regime labels. At deployment, the gate must distinguish recovery from tracking using four frames of proprioception alone. §III-B defines the recovery support Rδ as perturbed neighborhoods of get-up references, with Table V limiting reset perturbations to |v_lin|≤0.5 m/s, |ω|≤0.52 rad/s, joint noise ±0.1 rad. The push protocol (§IV-A) applies 525–575 N for 0.2 s, which will drive a G1 to velocities and contact configurations well beyond those bounds. Therefore the push-to-fall rollouts visit states that the paper explicitly excludes: 'states beyond this sampled support are outside our claim' (§III-B). The gate has never seen these states and its routing accuracy there is unknown; a misclassification either keeps the tracking expert active during a fall (chasing the infeasible command) or triggers recovery during ordinary tracking. The reported 100/100 is from a single deterministic run, with no error bars, so it does not establish distributional robustness. This is the load-bearing premise for the single-policy claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents StableMimic, a single proprioception-gated dual-expert policy for humanoid command tracking and post-fall recovery. Training augments the nominal tracking task with perturbed resets around human get-up references and a hidden successor-state reward; the deployed actor receives only the live command and proprioceptive history, with no get-up reference, recovery command, or external policy switch. Experiments on the Unitree G1 in MuJoCo report lower tracking errors on the LAFAN1 dance subset than a Single-MLP ablation and two prior systems, 100/100 recovery on a matched 100-push protocol, and qualitative real-hardware confirmation.","tokens_in":10633,"tokens_out":6000,"duration_ms":58292,"significance":"If the results hold, the paper makes a practically valuable contribution: it demonstrates a plausible path toward a single deployable policy that resumes uninterrupted command tracking after a fall, without external triggering or reference lookup, while improving post-fall motion and load metrics. The paper's strengths include a matched protocol across methods, an explicit training/deployment information-boundary table, a larger-capacity Single-MLP ablation, honest scoping statements, and careful differentiation between measured results and qualitative hardware demos. The central weakness is that the main claim about behavior during unplanned falls is evaluated under pushes that likely produce states outside the paper's declared recovery support, so the gate-transfer premise is not yet established.","major_comments":[{"comment":"The recovery support Rδ is explicitly an empirical set of perturbed neighborhoods around get-up references, with reset velocity bounds |v_lin|≤0.5 m/s, |ω|≤0.52 rad/s, and joint noise ±0.1 rad; the paper states 'states beyond this sampled support are outside our claim.' The push protocol applies 525–575 N for 0.2 s to a G1; a simple momentum estimate (Δv ≈ FΔt/m) already exceeds 0.5 m/s for plausible G1 masses, and the resulting contact configurations are unlikely to lie inside Rδ. Thus the main recovery evaluation exercises states the paper self-excludes. Because the gate in §III-C(a) was trained only on privileged rollout-regime labels within Rδ, its routing accuracy on push-induced states is unknown. The 100/100 result is from one deterministic protocol without error bars, so it cannot by itself establish robustness. Please report the empirical distribution of push-induced states rela","section":"§III-B / Table V / §IV-A"},{"comment":"Recovery success is defined only as 'a structured get-up and tracking resumption without resetting the policy or command.' The fall onset is quantified (pelvis height <0.50 m or tilt >60°), but 'structured get-up' is not. This makes the 100/100 success number non-reproducible and weakens the comparison across methods. Please specify an operational success criterion (e.g., pelvis height above a threshold, task tracking error below a bound for a duration, and no fall recurrence within the horizon).","section":"§IV-A(c)"},{"comment":"All comparisons are based on a single training run per method. RL policies have run-to-run variance; differences such as StableMimic vs Single-MLP MPBPE (28.53 vs 32.66 mm) or recovery success (100/100 vs 98/100) may not be significant. Without multiple seeds (or otherwise showing variance bounds), the claim that StableMimic achieves the lowest tracking and recovery metrics is unsupported in a statistical sense. Please report means and standard deviations (or distributions) over at least 3–5 seeds.","section":"Tables III and IV"}],"minor_comments":[{"comment":"Figure 4 is labeled 'conceptual' and the inset of Fig. 2 is qualitative. Consider adding a measured example of gate weights or routing probabilities over a fall-recovery rollout to support the 'soft blending' narrative.","section":"Fig. 4 and Fig. 2 inset"},{"comment":"The number of LAFAN1 dance sequences used for evaluation and the total duration of the 'complete dance subset' are not stated. Please specify these to improve reproducibility.","section":"§IV-A(a)"},{"comment":"The temperature τ in the softmax gate is not reported. Please state its value or scheduling.","section":"Eq. (3)"},{"comment":"The 'relative-height safeguard' is described only verbally. Please give the exact condition (e.g., h_reset_base ≥ 0.8 h_cmd) and clarify how it interacts with gate targets for non-reset transitions.","section":"§III-C(a)"},{"comment":"The reset ranges are reported per dimension; please clarify whether these are sampled uniformly and whether they apply to the root only or to all links/joints.","section":"Appendix Table V"}],"recommendation":"major_revision","confidential_remarks":"This is a solid experimental paper with a clear scope. My main concern is the gap between the training support and the push-protocol test distribution; this is fixable with additional analysis or retraining. I do not see grounds for rejection. Please also ask the authors to provide multiple seeds and a precise success criterion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"StableMimic is a solid, useful paper, and the main experimental claims are supported as far as the paper goes. The genuinely new thing is the single deployable policy that does both tracking and human-shaped post-fall recovery, with a proprioceptive gate that blends two experts and no runtime get-up reference, recovery command, or hard switch. That packaging removes a real failure mode of tracking-only policies, and the matched LAFAN1 tracking plus 100-trial recovery protocol is the right way to demonstrate it. I read the training curves and the recovery numbers as consistent with the intended expert separation; the Single-MLP comparison is fairer than most because it has a larger actor yet does worse. The authors are also honest about scope — they say states outside R_delta are outside the claim, and they call the hardware results qualitative. Credit where due.\n\nThe soft spots are real but not fatal. First, no code or data, and single deterministic runs. The tables have no seed variance or error bars, so 100/100 could be a point estimate from one lucky run; for an RL paper this is an addressable gap, not a flaw in the method. Second, the gate-transfer premise. The gate is trained with rollout-regime labels, and the push protocol (525–575 N for 0.2 s) likely visits states beyond the reset perturbations (linear velocity ±0.5 m/s, angular ±0.52 rad/s). The paper explicitly excludes such states. So the claim that the proprioceptive gate routes correctly outside R_delta is not demonstrated; 100/100 under one scripted protocol is thin evidence for a distributional claim. I would not call this circular — using get-up references in the reset/reward is standard reward shaping — but it is missing evidence. Third, the baselines are implemented by reference, which makes fairness hard to verify without code; the matched protocol helps, but only partly.\n\nWho is this for? Anyone working on humanoid tracking, teleoperation, or fall recovery in sim-to-real RL. It deserves a serious referee. My recommendation: send to peer review, and ask the authors to release code/data and rerun with seeds and error bars, plus a couple of disturbance protocols outside R_delta and gate-routing statistics. If those land, the paper is close to acceptance.","headline":"A genuinely useful unified tracking+recovery policy, with a real but addressable gap: the proprioceptive gate is only tested in-distribution and the stats are seedless.","tokens_in":11273,"tokens_out":2850,"would_cite":true,"duration_ms":31549,"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":"This paper claims that a single proprioception-gated policy can track humanoid motion, recover from falls in a human-like way, and resume the command without an external trigger or get-up reference.","keywords":["humanoid motion tracking","post-fall recovery","mixture-of-experts","proprioceptive gating","reinforcement learning","successor-state reward","get-up reference","Unitree G1"],"falsifier":"Run the deployed policy on falls whose initial states lie outside the sampled support used in training — for example, push directions not in {+x, -x, +y, -y}, or a fall from a small ledge — and record whether the recovery expert's gate weight stays high during ground contact and whether tracking resumes in the 1.5-s transition; a single such fall with sustained tracking-expert weight during low-height contact states would refute the claim that proprioception alone distinguishes recovery states from tracking states.","tokens_in":10227,"feed_emoji":"🤖","tokens_out":6415,"duration_ms":56036,"temperature":0.7,"pith_summary":"StableMimic tries to establish that a humanoid motion tracker can be trained to handle states outside its normal tracking distribution — specifically falls — within the same policy, so the robot recovers in a human-like, bounded way and resumes its command. It claims that perturbed training resets around human get-up demonstrations teach structured recovery, while a proprioception-only gate continuously blends a tracking expert and a recovery expert, avoiding a hard switch or runtime reference. If correct, one deployable policy can replace separate tracking and stand-up systems, and robots pushed over during dance or teleoperation will not thrash or chase unreachable commands. The paper reports the lowest errors on all four tracking metrics over complete LAFAN1 dance sequences among five methods, 100/100 recovery in matched push-to-fall trials with the lowest values on six of seven post-fall motion and load measures, and qualitative real-robot recovery on a Unitree G1.","feed_headline":"One policy tracks dance, falls, recovers, and resumes — no switch","feed_subtitle":"A proprioceptive gate blends a tracking and a recovery expert, so no external trigger or get-up reference is needed.","key_machinery":"The central object is a proprioception-gated dual-expert mixture-of-experts policy. Two experts map the same command-conditioned observation history to joint-target action means — one for tracking the live command, one for recovery shaped by hidden get-up references — while a soft gate takes only the proprioceptive history (base angular velocity, projected gravity, joint states, previous action) and computes a softmax blend of the two experts' outputs, so there is no discrete action switch. Training uses perturbed get-up resets around multiple human get-up trajectories to sample a recovery support beyond the command-tracking tube, and a hidden successor-state reward rewards approaching the n","core_discovery":"StableMimic's central claim is that extending a tracking policy's training support into perturbed neighborhoods of human get-up references, and routing post-fall states to a dedicated recovery expert through a proprioceptive gate, yields a single policy that both tracks a continuously advancing whole-body command and produces human-reference-shaped recovery without any runtime get-up reference, recovery command, or policy switch. In simulation, the gated dual-expert policy attains the best results on all four tracking metrics over the complete LAFAN1 dance subset, recovers in 100/100 matched push-to-fall trials, and achieves the lowest values on six of seven post-fall motion and load measure","pith_inferences":["Editorial extension: because the gate sees no command or reference, its ability to distinguish a fall from an intentionally low commanded pose is only as good as the four-frame proprioceptive history; a natural test is commanding a crouch profile and checking that tracking expert weight stays dominant.","Editorial extension: the recovery guarantee is explicitly bounded by the sampled perturbation neighborhoods around get-up references, so falls outside that support (new push directions, uneven terrain, ledges) may not recover in the demonstrated shape — a closed-loop curriculum that grows the support could be the next step.","Editorial extension: since reference identity and phase are hidden from the deployed Actor, the policy may be learning a general 'get upright' strategy rather than memorising specific get-up trajectories; evaluating on held-out get-up styles not present in the library would test this directly.","Editorial extension: the approach could be combined with fall-prediction or impact-mitigation policies: a predicted fall could trigger a protective pre-impact controller that hands over smoothly to the recovery expert via the same soft gate mechanism."],"forward_implications":["If correct, one deployable policy can track motion, fall, recover, and resume the command without external fall detection, a get-up reference, a recovery command, or a policy switch.","Training beyond the tracking distribution does not degrade tracking: the MoE policy achieves lower MPBPE and MJAE than a larger Single-MLP ablation and a tracking-only baseline on the complete LAFAN1 dance subset.","Post-fall motion and actuator load are reduced under the matched protocol: 100/100 recovery versus 98/100 for the Single-MLP and 0/100 for the tracking-only baseline, with lower limb speed, limb travel, joint speed, actuator torque, and positive energy.","The same exported ONNX policy runs on real Unitree G1 hardware at 50 Hz, and both dance and constant-standing-reference deployments show bounded limb motion, autonomous recovery, and task resumption.","Because the method uses a common observation/action interface, it can augment other RL trackers with compatible spaces, extending structured post-fall behavior to teleoperation-oriented systems."],"fun_headline_variants":["Dance, fall, recover: one policy, no switch — StableMimic","Humanoid learns to get up like a human, no external trigger","Tracking plus recovery in one policy: 100/100 fall recoveries","Proprioceptive gate blends track and recover experts seamlessly"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that a four-frame history of the robot's own body readings alone is enough, at deployment, to tell when the robot is in a post-fall recovery state rather than still tracking a commanded pose, because the gate's routing labels are available only during training and the command/reference is never given to it at runtime; the paper itself limits this claim to the sampled perturbation neighborhoods around get-up references.","fun_headline_variants_meta":{"raw":{"variants":["Dance, fall, recover: one policy, no switch — StableMimic","Humanoid learns to get up like a human, no external trigger","Tracking plus recovery in one policy: 100/100 fall recoveries","Proprioceptive gate blends track and recover experts seamlessly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"total_tokens":1214,"prompt_tokens":806,"completion_tokens":408,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":330}},"tokens_in":550,"tokens_out":408,"duration_ms":53185,"temperature":1.0,"reasoning_tokens":330,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T08:13:19.539151+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the deployed policy on falls whose initial states lie outside the sampled support used in training — for example, push directions not in {+x, -x, +y, -y}, or a fall from a small ledge — and record whether the recovery expert's gate weight stays high during ground contact and whether tracking resumes in the 1.5-s transition; a single such fall with sustained tracking-expert weight during low-height contact states would refute the claim that proprioception alone distinguishes recovery states from tracking states.","supporting_citations":[],"review_version":1}