{"id":"cc763214-ad1c-43e3-9e87-24e0ed1adcdc","arxiv_id":"2608.07328","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A learned controller with latent alignment and a learnable gait-frequency action keeps a 68 kg quadruped moving after sudden actuator power loss, in simulation and on hardware.","lead":"The authors train a 68 kg quadruped robot to keep walking after a leg actuator loses power, using reinforcement learning in simulation. The controller learns to change step timing to adapt to the fault, and it transfers to the real robot on flat ground.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The learnable gait-frequency ablation is confounded: the phase-reward target is a function of the policy's own frequency action, so the claimed benefit of adaptive gait timing is not causally demonstrated.","rationale":"The reader identified the sim-to-real fault model as the weakest assumption. That is a valid concern about validation breadth, but the more load-bearing issue is internal to the method's central claim: the evidence for the 'adaptive gait timing' mechanism is confounded. The phase reward is constructed from a reference phase that the policy itself controls via aν, creating a self-referential reward. The free-gait baseline in Section IV-C.2 removes aν and changes the reward from feet-phase to air-time simultaneously, so the comparison cannot isolate the frequency action's contribution. If the frequency action is not actually responsible for the improved gait stability, the paper's primary novelty ('augment the action space with a learnable gait frequency parameter, enabling adaptive gait timing') is unsupported, even though the overall fault-tolerant policy may still work. This concern can be resolved by a targeted ablation that holds the reward fixed and varies only whether the phase reference uses the learned frequency or a constant. The paper's other ablations (latent alignment, history length) are cleaner, and the real-robot demonstration is a useful existence proof, but the gait-frequency contribution requires this additional experiment before the central claim is established.","tokens_in":73,"tokens_out":11910,"duration_ms":177499,"concrete_test":"Train a fixed-frequency variant that keeps the same feet-phase reward but sets νref to a constant ν_def (no aν in the reward path, while optionally keeping aν in the action space), and compare its tracking/survival metrics to the full method across the same fault suite. If the fixed-frequency variant performs similarly, the frequency action is not the source of the improvement. Also check whether the learned aν tracks the actual contact transitions; if it does, the phase reward is being gamed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim includes 'adaptive gait timing through a learnable frequency action' as a key mechanism. In Section III-B, the reference phase is updated using νref = ν_def + sν·aν, and in Section III-C the feet-phase reward (Tab. II, Eq. 2) compares measured contacts to a reference contact schedule cref(φ) derived from this same phase. Thus the reward target is endogenous: the policy can increase the phase reward by adjusting aν to shift cref toward whatever contact pattern it already produces, rather than by coordinating to a fixed temporal reference. This undermines the interpretation of aν as causing adaptive gait timing. The ablation in Section IV-C.2 changes two variables at once: it removes the frequency action and replaces the feet-phase reward with an air-time reward. Any performance difference between the full method and this 'free-gait' baseline could come from the reward change alone, or from reward gaming via the manipulable phase, rather than from the frequency action. Because the paper's stated novelty depends on the frequency action being beneficial, this confound directly weakens the support for the central claim.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a reinforcement-learning controller for a 68 kg quadruped (Kyon) that maintains locomotion after a sudden actuator power-loss fault. The architecture is an asymmetric actor-critic in which the critic uses privileged fault information while the actor is trained with a latent-alignment loss to reconstruct that information from a history of proprioceptive observations. The action space is augmented with a scalar gait-frequency action aν that modulates the reference phase used by a feet-phase reward. Training is done with PPO in MJX on stepped-pyramid terrain with a torque-efficiency curriculum, and evaluation includes five-seed learning curves, simulation tests with random faults, a sim-to-sim MuJoCo/XBot2 study, and one real-hardware flat-ground experiment with a rear-left knee fault.","tokens_in":12522,"tokens_out":5225,"duration_ms":54765,"significance":"If the mechanism claims are supportable, the paper contributes a useful single-policy approach to fault-tolerant locomotion: the policy is fault-unaware at deployment, no predefined faulty-leg strategy is prescribed, and the hardware result on a 68 kg quadruped is a meaningful step beyond small-platform demonstrations. Strengths of the manuscript include five-run confidence intervals on the learning curves, detailed reporting of network sizes and PPO hyperparameters, a realistic sim-to-sim setup with asynchronous XBot2 communication, and an honest distinction between uneven-terrain simulation and flat-ground hardware validation. The main weakness is that the headline 'adaptive gait timing' benefit is not causally identified by the presented ablation, because the reference phase used in the reward is a function of the policy's own frequency action and because the ablation changes the reward function at the same time as the action space.","major_comments":[{"comment":"The feet-phase reward is endogenous in a way that undermines the causal interpretation of the learnable gait-frequency action. In Section III-B the reference frequency is νref_t = ν_def + sν aν_t, and the same phase update is used to generate the reference contact schedule cref in Eq. (2) and Table II. Thus the policy can increase its feet-phase reward by moving aν_t so that cref shifts toward the contact pattern the policy already produces, rather than by coordinating to a fixed temporal reference. The paper's stated novelty depends on aν being beneficial as an adaptation mechanism, so this endogeneity weakens the support for that central claim.","section":"III-B, III-C, Eq. (2)"},{"comment":"The learnable-gait-frequency ablation is confounded because it changes two variables at once. The text states that, 'For a fair comparison, we replace the feet-phase reward term ... with the widely used feet air-time reward.' The full method therefore differs from the 'free-gait' baseline both in the presence of aν and in the reward function, and any observed performance difference could come from the air-time reward alone rather than from the frequency action. A clean ablation should keep the reward fixed (for example, a feet-phase reward computed from the default frequency for both policies) and remove only the frequency action from the action space.","section":"IV-C.2, Fig. 9, Fig. 10"},{"comment":"The main simulation performance metrics in Fig. 5 are reported without error bars or confidence intervals. The figure is the primary deployment-time evidence that the full method outperforms its ablations across fault locations, yet it appears to be based on one selected policy per training run without reporting variance over the five runs or over the 1024 test agents. The learning curves in Fig. 4 already use 95% confidence intervals, and the same standard should be applied to the evaluation metrics so that the comparison is statistically grounded.","section":"IV-B, Fig. 5"}],"minor_comments":[{"comment":"The real-hardware validation covers a single fault configuration (rear-left knee pitch joint) on flat ground with terrain observations set to zero. This is a useful proof of concept, but the text could state more prominently that hardware transfer to other joints, partial torque degradation, and uneven terrain remains untested.","section":"IV-B.2"},{"comment":"The fault model is a scalar torque-efficiency scaling τ_j ← kτ,j τ_j applied to the PD output. The paper does not discuss how real failures involving partial torque collapse, increased joint friction, or communication latency would map to this model; a brief limitation sentence would clarify the intended scope of the sim-to-real claim.","section":"III-D, Table V"},{"comment":"The caption says the method 'outperforms the same architecture without proprioceptive history (green) and without latent-alignment loss (yellow),' but the comparison with the oracle is also shown; it would be clearer to state explicitly that the red curve is closest to the oracle among the non-privileged variants.","section":"Fig. 4 caption"},{"comment":"The relationship between the 25 s experiment duration and the reported maximum survival time of 20 s is confusing; since the fault occurs at t = 5 s, the maximum post-fault survival time is 20 s, and this should be stated in the text to avoid an apparent inconsistency.","section":"IV-B, Fig. 5"}],"recommendation":"major_revision","confidential_remarks":"The central issue is identification of the gait-frequency mechanism, not the correctness of the training pipeline. If a reward-matched ablation shows a genuine benefit of aν, the paper would be much stronger; if it does not, the 'adaptive gait timing' contribution should be softened. The variance reporting in Fig. 5 should also be addressed before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honestly, this is a useful paper but the central claim about adaptive gait timing is less supported than it looks. The method—asymmetric actor-critic with a latent-alignment loss, a learnable gait frequency action, and a curriculum over torque efficiency—is a sensible combination, and the hardware demo on a 68 kg quadruped surviving a knee power loss is a real result. Credit where due: they ran five training seeds, show confidence intervals on learning curves, and did sim-to-sim with the XBot2 middleware, which is more than many locomotion papers do.\n\nThe soft spots are real but not fatal. The main tracking metrics in Fig. 5 have no error bars, so you can't tell if the reported per-joint differences are meaningful. The hardware validation is one scenario on flat ground—encouraging but thin. And there's no comparison to existing fault-tolerant RL methods like FT-Net or DreamFlex, which limits the strength of the novelty claim.\n\nThe biggest issue is the gait-frequency ablation. The baseline removes the frequency action and replaces the feet-phase reward with an air-time reward. That's two changes at once, so the comparison can't isolate the frequency benefit. Worse, the phase reward is endogenous: the reference phase is computed from the policy's own frequency action, so the policy can shift the reference schedule toward the contacts it already produces. That makes the reward easier to satisfy without any real temporal coordination. This doesn't sink the whole paper—the overall controller clearly works—but it means the 'adaptive gait timing' claim is not causally demonstrated. The authors need a clean ablation that holds the reward fixed and only removes the frequency action, or at least an analysis that accounts for the endogeneity.\n\nThe method itself is reasonably novel as a combination, and the fault-inference analysis with history length is a nice empirical finding. The paper is worth serious peer review; it deserves a revision, not a desk reject. Send it back and make them fix the ablation, add error bars, and ideally a second hardware scenario.","headline":"A competent fault-tolerant locomotion paper with a credible hardware demo, but the adaptive gait timing claim rests on a confounded ablation and needs a cleaner experiment.","tokens_in":13038,"tokens_out":2640,"would_cite":true,"duration_ms":24379,"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":"A single policy keeps a 68 kg quadruped walking after sudden joint power loss.","keywords":["fault-tolerant locomotion","quadruped robots","reinforcement learning","asymmetric actor-critic","latent alignment","gait frequency adaptation","power loss fault","sim-to-real transfer"],"falsifier":"Compare the trained policy's survival time and command-tracking error on the physical robot under a partial torque collapse (for example $k_{\\tau,j} = 0.5$) and a hip-joint fault during stair descent, against the simulation values reported in Fig. 5; if real-robot performance falls far below the simulated distribution, the zero-shot sim-to-real claim and the fault model are contradicted.","tokens_in":12050,"feed_emoji":"🤖","tokens_out":9611,"duration_ms":70900,"temperature":0.7,"pith_summary":"This paper tries to show that a single control policy, trained once in simulation, can keep a 68 kg quadruped walking after one of its actuators suddenly loses power, without being told which leg is damaged or how to alter the gait. The robot infers the fault from a short history of its own joint readings and copes by learning to change step timing through a gait-frequency action, rather than by adopting a predefined tripod or three-legged strategy. The authors argue that such adaptive timing is the right response for heavier robots, where the mass and actuation limits make the fast corrective movements used on smaller quadrupeds impractical. They back this with simulation results on uneven terrain, ablations showing the value of history and of the alignment loss, and a real-robot demonstration on flat ground after a knee joint fails. If the claim holds, it would mean that robust fault tolerance can be bought with a single learned controller rather than a library of failure-specific responses.","feed_headline":"One policy keeps a 68 kg robot walking after a joint fails","feed_subtitle":"Trained once in simulation, the controller infers the failed joint from sensor history and retimes its gait steps.","key_machinery":"The central machinery is an asymmetric actor-critic architecture reinforced by two additions on top of the standard PPO objective. First, a latent-alignment loss $\\mathcal{L}_{\\mathrm{MSE}} = \\mathbb{E}[(\\hat{r}_t - r_t)^2]$ forces the actor encoder's latent representation $\\hat{r}_t$, computed from a history $h_t = \\langle o_t,\\ldots,o_{t-H+1}\\rangle$ of proprioceptive observations, to match the critic encoder's latent $r_t$, which is computed from privileged observations including the joint fault mask $m_{J,t}$. Second, the action space is augmented with a scalar gait-frequency action $a^\\nu_t$ that modulates the reference stepping frequency $\\nu^{\\mathrm{ref}}_t = \\nu^{\\mathrm{def}} + s_\\nu a^\\nu_t$, whose phase update $\\phi_{t+1,\\ell} = \\mathrm{mod}(\\phi_{t,\\ell} + 2\\pi\\Delta t\\,\\nu^{\\mathrm{ref}}_t + \\pi, 2\\pi) - \\pi$ drives the reference contact schedule used in the reward. Power loss is encoded in training by scaling the PD torque output as $\\tau_j \\leftarrow k_{\\tau,j}\\tau_j$ with a curriculum over $k_{\\tau,j}$, and faulty legs are excluded from the feet-phase reward so the policy is never pushed toward a specific compensation gait.","core_discovery":"On its own terms, the paper's central claim is that an asymmetric actor-critic reinforcement learning setup, augmented with a latent-alignment loss and a learnable gait-frequency action, yields a single policy that maintains locomotion under sudden actuator power loss on a 68 kg quadruped, with no predefined faulty-leg strategy. The critic is trained with privileged information that explicitly includes a joint fault mask, while the deployed actor sees only a history of noisy proprioceptive observations plus terrain height; a mean-squared-error term forces the actor's latent code toward the critic's privileged code, and the phase-consistency reward is masked for faulty legs so the policy is free to discover contact schedules that work with a dead joint. Torque loss is modeled as scaling the commanded torque by a factor $k_{\\tau,j} \\in [0,1]$ on the failed joint, with a curriculum that gradually deepens the fault severity based on velocity-tracking performance. The result is a policy whose gait-frequency action rises during single support and falls during double support after a knee fault, effectively switching to a tripodal pattern with longer, more periodic stance phases, and which transfers zero-shot to the real Kyon robot for a rear-left knee-pitch power loss on flat ground.","pith_inferences":["Beyond the paper's experiments, the latent-alignment loss likely also transfers non-fault privileged information (terrain, dynamics); a healthy-only ablation would reveal whether alignment still helps when no fault is present, which the paper does not report.","Because the real-world validation uses flat ground with terrain observations zeroed, the claims about uneven-terrain deployment depend on the perception pipeline; testing on uneven real terrain with onboard terrain reconstruction is a direct next step the paper lists but does not take.","The gait-frequency action's periodic modulation suggests the policy learns an internal phase clock tied to contact state; one testable implication is that commanding a different nominal frequency $\\nu^{\\mathrm{def}}$ would preserve the ratio of single-to-double support durations, which the paper does not measure.","A sharper falsification would target hip-roll faults during stair descent, where the passively swinging leg's dynamics differ most from the tested knee-pitch case; the paper's survival-time plots already show hip-roll faults are among the harder cases in simulation."],"forward_implications":["A single policy, trained once in simulation, can be deployed without any fault-detection or leg-status input; the fault is inferred from proprioceptive history alone.","Under knee-joint faults the policy spontaneously adopts a tripodal gait with longer, more periodic stance phases, and the learned gait-frequency action rises in single support and falls in double support.","The controller generalizes in simulation to novel terrain profiles (stairs with 10 cm steps and ramps up to 13 degrees) without retraining.","Increasing the proprioceptive history beyond two steps yields only marginal gains, indicating that one-step temporal differences carry most of the information needed to reconstruct the privileged fault state.","For heavier quadrupeds, adaptive step timing may be the scalable fault-response strategy, because the robot's mass and actuation limits preclude the aggressive high-frequency corrections used on smaller platforms."],"supporting_citations":[{"why":"Supplies the asymmetric actor-critic training scheme that lets the critic see privileged information while the actor only sees sensor observations.","marker":"[19]"},{"why":"Provides the PPO objective used as the base learning algorithm for all variants.","marker":"[35]"},{"why":"Provides the MuJoCo physics engine used in the sim-to-sim evaluation and in the deployment-style test with XBot2.","marker":"[38]"},{"why":"Provides the XBot2 middleware that reproduces asynchronous 1 kHz communication in simulation, making the sim-to-real setup more realistic.","marker":"[39]"},{"why":"The progressive fault-severity curriculum (increasing torque loss based on tracking performance) is adopted from this prior work.","marker":"[32]"},{"why":"Describes the Kyon 68 kg quadruped platform on which the real-robot experiments are run.","marker":"[5]"},{"why":"MuJoCo Playground supplies the locomotion reward terms that the paper's reward function is based on.","marker":"[36]"}],"fun_headline_variants":["Adaptive gait timing keeps 68kg robot walking after joint fails","Single policy retimes steps to handle actuator loss on quad","Robot learns to walk with a dead joint via adaptive gait timing","One policy, no failed-leg strategies: adaptive gait timing works","Trained once, this policy infers failed joint and retimes steps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's real-world validity rests on the assumption that the torque-scaling model of power loss used in training, with its domain-randomization ranges, reproduces real actuator failures closely enough that a policy trained in simulation works zero-shot on the physical robot; the paper only tests this for one joint at one severity on flat ground.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive gait timing keeps 68kg robot walking after joint fails","Single policy retimes steps to handle actuator loss on quad","Robot learns to walk with a dead joint via adaptive gait timing","One policy, no failed-leg strategies: adaptive gait timing works","Trained once, this policy infers failed joint and retimes steps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000182,"raw_usage":{"total_tokens":1312,"prompt_tokens":951,"completion_tokens":361,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":273}},"tokens_in":567,"tokens_out":361,"duration_ms":3673,"temperature":1.0,"reasoning_tokens":273,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T10:11:56.629041+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the trained policy's survival time and command-tracking error on the physical robot under a partial torque collapse (for example $k_{\\tau,j} = 0.5$) and a hip-joint fault during stair descent, against the simulation values reported in Fig. 5; if real-robot performance falls far below the simulated distribution, the zero-shot sim-to-real claim and the fault model are contradicted.","supporting_citations":[{"cited_title":"The xbot2 real-time middleware for robotics,","cited_arxiv_id":null,"evidence_quote":"Provides the XBot2 middleware that reproduces asynchronous 1 kHz communication in simulation, making the sim-to-real setup more realistic."},{"cited_title":"Dreamflex: Learning fault-aware quadrupedal locomotion controller for anomaly situation in rough terrains,","cited_arxiv_id":null,"evidence_quote":"The progressive fault-severity curriculum (increasing torque loss based on tracking performance) is adopted from this prior work."},{"cited_title":"KYON: Semi-Modular Wheel-Legged Quadruped With Agile Bimanual Capability","cited_arxiv_id":"2606.30243","evidence_quote":"Describes the Kyon 68 kg quadruped platform on which the real-robot experiments are run."},{"cited_title":"Mujoco playground: An open-source framework for gpu-accelerated robot learning and sim-to-real transfer","cited_arxiv_id":null,"evidence_quote":"MuJoCo Playground supplies the locomotion reward terms that the paper's reward function is based on."}],"review_version":1}