{"id":"cbc8f61f-4e7f-41d3-ac70-02658d4a1b62","arxiv_id":"2509.18046","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A single-layer Mamba encoder as the policy backbone improves learning speed, stability, and energy efficiency of an end-to-end RL humanoid walking controller in simulation compared to a feedforward baseline.","lead":"The paper applies a Mamba state-space encoder instead of a feedforward network as the backbone of a reinforcement-learning controller for a simulated humanoid robot walking on planned footsteps. Across five walking and standing tasks, the Mamba version learns faster, walks with lower torque and power use, and reaches higher rewards than the feedforward baseline.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Energy/torque gains may reflect slower gait, not Mamba: no speed control in reward or evaluation, and Table VI implies HuMam is ~42% slower on curved walking.","rationale":"The reader's weakest assumption was the oracle footstep planner and the 'end-to-end' claim. That is a real scope limitation, but it does not threaten the core quantitative comparison between HuMam and the feedforward baseline under identical oracle inputs. The more load-bearing risk is that the measured energy and torque advantages may be artifacts of a confound in the task definition: the reward never rewards speed, and the evaluation does not control for it. The derived speed mismatch in curved walking is concrete and testable. If the policies differ in speed, all downstream energy and torque numbers are not interpretable as 'Mamba improves control economy.' This concern is distinct from the reader's, hence 'disagree.' I still recommend the same overall verdict (CONDITIONAL) because the learning-efficiency and return claims are less affected, and the speed confound could be remedied by a controlled experiment; the paper should not be rejected outright, but its energy-saving claims need an explicit speed-matched check.","tokens_in":11877,"tokens_out":8837,"duration_ms":81891,"concrete_test":"Re-run or re-analyze the saved rollouts used for Tables V and VI: (1) compute mean root speed per task for both policies; (2) confirm the implied curved-walking speeds (HuMam ≈ 0.073 m/s vs baseline ≈ 0.125 m/s); (3) retrain or re-evaluate with a speed-tracking reward, or filter rollouts to matching speed bins, and recompute J/m, W, W/kg, and torque peaks. If the energy and torque advantages disappear at matched speeds, the headline energy claims are confounded by gait speed rather than caused by the Mamba backbone.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing issue is the absence of speed control in the task definition and evaluation. The reward (Eq. 5) contains no velocity or progress term, and footstep advancement is based on holding a foot inside a 20 cm radius of the target for one single-support duration (Sec. IV-B.c), so a slower gait is never penalized. In the energy metrics (Sec. IV-C, Table VI), speed is not reported, but it can be recovered from power divided by energy per meter (W / (J/m)). For curved walking, HuMam's derived speed is 95.17/1303 ≈ 0.073 m/s versus the baseline's 131.93/1055 ≈ 0.125 m/s, a 42% speed deficit. Since joint torques and actuation power scale with gait speed, the reported reductions in power, energy, and torque peaks may simply be consequences of walking slower, not of the Mamba encoder. This confound directly undercuts the abstract's 'reducing power consumption and torque peaks' and the conclusion's 'energy saving' claims. Even in forward walking, where speeds are closer, no speed-matched controlled evaluation is provided, so the torque improvements in Table V remain potentially speed-related.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents HuMam, an end-to-end RL framework for humanoid locomotion that uses a single-layer Mamba encoder to fuse robot-centric proprioceptive states with two planned footstep targets and a continuous phase clock. The policy outputs 12 joint-position targets tracked by a PD controller and is trained with PPO under a six-term reward. Experiments are conducted in mc-mujoco on the JVRC-1 humanoid across forward, backward, lateral, curved walking, and standing, comparing HuMam against a feedforward baseline with identical training settings. The central claims are that the Mamba backbone improves learning efficiency, training stability, and final returns, while also reducing power consumption and torque peaks.","tokens_in":12226,"tokens_out":3965,"duration_ms":32119,"significance":"If the empirical claims hold, the paper makes a modest but useful contribution: it is the first demonstration of a Mamba-style state-space encoder as the fusion backbone for end-to-end humanoid locomotion RL, and it provides a systematic comparison against a strong feedforward baseline. The paper is clearly written, the method is straightforward, and the training/evaluation setup is shared across both arms of the comparison, which limits circularity. However, the absence of speed control and speed reporting in the energy/torque evaluations, combined with missing seed counts and statistical tests, currently weakens the quantitative support for the headline contributions. The work does not provide code, checkpoints, or hardware validation, so the claims rest entirely on simulation-based comparisons.","major_comments":[{"comment":"Speed confound in energy/torque claims: The reward in Eq. (5) contains no velocity or progress term, and step achievement (§IV-B.c) only requires a foot to remain within 20 cm of the target for one single-support duration. Recovering walking speed from Table VI as power ÷ (J/m) gives, for curved walking, HuMam at 95.17/1303 ≈ 0.073 m/s versus Baseline at 131.93/1055 ≈ 0.125 m/s—a 42% speed deficit. Forward walking is also slower (≈0.256 vs 0.272 m/s). Since joint torques and actuation power scale with gait speed, the reported reductions in power, energy per meter, and torque peaks may simply reflect a slower gait rather than the Mamba encoder. The authors must report walking speed for every task and provide a speed-matched controlled evaluation (e.g., matched commanded speed or evaluation at the same achieved speed) before claiming 'reducing power consumption and torque peaks'.","section":"§IV-C, Table VI, Eq. (5)"},{"comment":"Lack of statistical support: The paper reports learning curves 'across seeds' and standard deviations, but never states the number of seeds. No statistical significance tests or confidence intervals are provided for the small differences in Table IV (e.g., average total reward 0.737 vs 0.728). Moreover, for curved and forward walking the Baseline actually obtains higher total reward (0.677 vs 0.672 and 0.694 vs 0.668, respectively), which contradicts the unqualified 'consistently improves' claim. Add the seed count, per-seed results, and significance tests or effect sizes; without these, the central comparison is not rigorously established.","section":"§IV-D, Tables III–IV, Fig. 3"},{"comment":"Torque comparison is not speed-matched: Table V reports average and peak torques only for forward walking, and no walking speed is given. Given the speed differences inferred from Table VI, the reported 9.6% average and 9.1% peak torque reductions could be consequences of slower gait. The authors should either report torque at matched gait speeds or include speed as a covariate; otherwise the 'smoother actuation' and 'energy-saving' conclusions are not supported.","section":"§IV-C, Table V"},{"comment":"Domain randomization is unspecified: Algorithm 1 samples domain-randomization parameters ξ, and the text mentions dynamics, sensors, init state, and target jitter, but the actual distributions and ranges are never provided. This is a reproducibility gap and also weakens the robustness interpretation. Provide the full DR specification (e.g., noise magnitudes, friction ranges, mass distributions) or state explicitly that DR was not used.","section":"Algorithm 1, §IV-B.d"}],"minor_comments":[{"comment":"Energy metrics are reported without error bars or confidence intervals, even though the paper acknowledges variability elsewhere. Add error bars for J/m, W, and W/kg, or explain why they are omitted.","section":"Table VI"},{"comment":"The phrase 'one single-support duration' is used to define when a step is scored but is never defined. Clarify how this duration is determined and how target advancement is triggered.","section":"§IV-B.c"},{"comment":"The state-space encoder equations use σ(·) for 'elementwise gates', but standard Mamba uses specific nonlinearities (e.g., SiLU, softmax for selection). Clarify the exact gating functions and tensor dimensions of W_A, W_B, W_C, and W_D, since the equations as written are not self-contained.","section":"§III-C, Eqs. (15)–(16)"},{"comment":"In the text, the right hip pitch (R HIP P) is described as showing a 'substantial reduction in both average and peak torque', but the table shows average −16.4% and peak +41.1%. Verify the reported numbers or rephrase; the current wording is misleading for that joint.","section":"Table V"},{"comment":"The caption says 'across scenarios', but it is not clear whether the learning curves are aggregated over all tasks or shown per task. Label the panels or explain the aggregation.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The speed confound is the load-bearing issue: if the authors can provide speed-matched energy/torque comparisons and seed-level statistics, the empirical claims would be substantially strengthened. The manuscript is within the scope of the journal, but the current evidence is insufficient for the strong claims made in the abstract and conclusion. The lack of code or checkpoints also limits reproducibility; I would encourage the authors to release them."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my take on arXiv:2509.18046. The genuinely new thing is that a single-layer Mamba encoder can serve as the fusion backbone for a humanoid walking policy trained with PPO, and the paper shows a concrete, reproducible-in-principle empirical gain over a feedforward baseline: it reaches return thresholds with 13–42% fewer samples and shows lower cross-seed variance. That part looks credible, because the two policies are trained with identical rewards and hyperparameters, and the learning curves are consistent.\n\nWhere it gets shaky is the energy and torque claims. The reward has no velocity or progress term, and footstep scoring only requires a foot to stay near a target for one single-support duration, so the policy is never penalized for walking slowly. Recovering speed from Table VI: on curved walking, HuMam's derived speed is about 0.073 m/s versus 0.125 m/s for the baseline, a 42% deficit. HuMam actually consumes more energy per meter on that task (1303 vs 1055 J/m), yet the text spins it as 'compensating with smoother dynamics and reduced power-per-mass.' Power-per-mass is just average power divided by body mass, so of course it drops when you walk slower. The forward walking energy numbers also show a small speed deficit, so the torque reductions in Table V could be partly gait-speed effects too. No speed-matched evaluation is provided, so the abstract's 'reducing power consumption and torque peaks' claim is not established.\n\nMinor issues: seed counts are not reported, no statistical significance tests, domain randomization and PD gains are unspecified, and no code or data are given. The 'end-to-end' label is a stretch because the policy consumes externally planned footstep targets as oracle inputs, with online planning left to future work.\n\nNet: the learning-efficiency and stability result is likely real and worth refereeing, but the energy-saving claims need to be reworked with speed-controlled evaluation and honest reporting of the curved-task result. I'd send it to a serious referee, but with the expectation of substantial revision.","headline":"Mamba encoder shows real learning-efficiency gains for humanoid walking, but the energy-saving claims are confounded by slower gait and need a speed-matched re-evaluation.","tokens_in":12684,"tokens_out":3554,"would_cite":false,"duration_ms":28833,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Replacing the policy network with a Mamba encoder makes humanoid walking RL faster, stabler, and cheaper to actuate.","keywords":["humanoid locomotion","deep reinforcement learning","Mamba","state-space models","footstep planning","energy efficiency","PPO","JVRC-1"],"falsifier":"Train the same Mamba and feedforward policies with footstep targets randomly jittered or drawn from a distribution not seen in training, and measure the return gap; if the feedforward baseline matches or beats Mamba under target noise, the claimed advantage is not about feature fusion but about overfitting to clean plans. A more direct test: replace the Mamba layer with a linear layer of equal parameter count and check whether the sample-efficiency and torque-reduction gains persist.","tokens_in":11802,"feed_emoji":"🤖","tokens_out":2723,"duration_ms":29053,"temperature":0.7,"pith_summary":"The paper argues that a single-layer Mamba encoder—a selective state-space model—can serve as the fusion backbone for end-to-end reinforcement learning of humanoid locomotion. On the simulated JVRC-1 robot, this backbone learns to walk forward, backward, sideways, along curves, and stand in place, using a policy that maps robot states and two planned footsteps to joint targets. Compared with a feedforward network trained identically, the Mamba policy reaches target returns in up to 42.5% fewer samples, shows 35–61% lower training variance, and reduces both average and peak joint torques and power consumption. The authors claim this is the first end-to-end humanoid RL controller using Mamba as the fusion backbone.","feed_headline":"Mamba encoder speeds humanoid RL and cuts torque","feed_subtitle":"Single-layer state-space backbone beats feedforward net in sample efficiency, stability, and energy use on JVRC-1.","key_machinery":"The single-layer Mamba encoder with selective gating (state-space update equations) that mixes two tokens: robot-centric state (leg joint positions/velocities, base orientation/angular velocity) and external state (two oriented footstep targets plus a continuous phase clock). The encoder's output feeds shared policy and value heads optimized with PPO; actions are joint position targets executed by a low-gain PD loop at 1000 Hz. A six-term reward shapes contact, swing, step placement, posture, height, and upper-body stability.","core_discovery":"The central claim is that a state-centric Mamba encoder—processing a token pair of robot-centric and external footstep states through gated state-space dynamics—yields more efficient, stable, and physically economical humanoid walking policies than a feedforward baseline trained under identical settings. The discovery is an application result: the Mamba structure, not new rewards or simulators, is credited for faster convergence, lower cross-seed variance, and reduced actuation cost across forward, backward, lateral, curved, and standing tasks.","pith_inferences":["The observed benefits may stem from Mamba's selective gating acting as an input-dependent feature mixer, effectively learning to weight robot versus footstep information per state; a controlled ablation replacing Mamba with a single gated linear layer of equal parameter count would test this mechanism directly.","If the advantage persists under noisy or partially occluded footstep targets, the encoder likely learns robust representations rather than memorizing exact plan coordinates; this could be checked by perturbing target positions at evaluation time.","The same backbone may benefit other contact-rich legged robots or whole-body manipulation tasks with multimodal state inputs, though this is untested and speculative.","Because the paper only compares against one feedforward baseline, the relative benefit over other sequence models (GRU, Transformer) under the same compute envelope remains open; the authors themselves list this comparison as future work."],"forward_implications":["If the Mamba advantage holds, a lightweight single-layer state-space encoder could replace heavier recurrent or attention-based backbones in onboard humanoid controllers with limited compute and memory.","The reported 9.6% mean and 9.1% peak torque reductions, plus lower energy per meter, could translate to longer battery life and reduced mechanical stress on physical robots.","The state-centric interface (robot states plus two planned footsteps) is compatible with adding perception later without changing the policy's input structure.","The sample-efficiency gains (up to 42.5% fewer samples to reach a target return) could reduce training cost for new gaits or terrains.","The consistent improvement across distinct walking modes suggests the Mamba backbone generalizes across gait patterns, not just one trajectory type."],"fun_headline_variants":["Mamba backbone stabilizes humanoid RL, lowers torque peaks","State-centric Mamba cuts energy, boosts RL stability","First Mamba-based humanoid RL: efficient, stable, economical","Mamba fusion beats feedforward in humanoid RL control"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The policy assumes the two upcoming footstep targets it receives are correct and always available; the Mamba backbone's demonstrated gains apply only under these oracle plans, with online footstep planning and feasibility checks deferred to future work.","fun_headline_variants_meta":{"raw":{"variants":["Mamba backbone stabilizes humanoid RL, lowers torque peaks","State-centric Mamba cuts energy, boosts RL stability","First Mamba-based humanoid RL: efficient, stable, economical","Mamba fusion beats feedforward in humanoid RL control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000249,"raw_usage":{"total_tokens":1350,"prompt_tokens":673,"completion_tokens":677,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":417,"completion_tokens_details":{"reasoning_tokens":608}},"tokens_in":417,"tokens_out":677,"duration_ms":6441,"temperature":1.0,"reasoning_tokens":608,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T15:46:27.606857+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same Mamba and feedforward policies with footstep targets randomly jittered or drawn from a distribution not seen in training, and measure the return gap; if the feedforward baseline matches or beats Mamba under target noise, the claimed advantage is not about feature fusion but about overfitting to clean plans. A more direct test: replace the Mamba layer with a linear layer of equal parameter count and check whether the sample-efficiency and torque-reduction gains persist.","supporting_citations":[],"review_version":1}