{"id":"9ce975b0-aafc-4c69-b90f-089d10dc8cfd","arxiv_id":"2502.05817","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DreamFLEX uses an explicit fault-estimation network to detect broken leg joints and modulate a learned locomotion policy, letting a quadruped traverse rough terrain with fewer working legs.","lead":"DreamFLEX is a robot controller that helps four-legged robots keep walking over rough terrain even when one of their joints is broken or a motor gets weak. It works by having the robot estimate which joints are damaged and change its step pattern to avoid using them.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The weakened-motor claim is untested: evaluation only uses kτ=0.0 (full disablement), and the binary fault vector (Eq. 5) cannot represent partial torque loss, so robustness to genuinely weakened motors is not demonstrated.","rationale":"After careful reading, I find the principal risk to the central claim is not the overall architecture but the mismatch between the fault representation/evaluation and the claimed failure mode. DreamFLEX's explicit contribution is a binary joint fault vector (Eq. 5) and a modulation layer (Eq. 9) that conditions the policy on it. The paper's own training procedure (Section II.B.2) samples kτ continuously, but the evaluation (Section III.B) fixes kτ=0.0, which is complete motor disablement, not weakening. Thus the central claim about 'weakened motors' is not directly tested. A binary f_t cannot encode the continuous severity of a partial torque loss; the policy must infer severity solely from the observation history through the latent z_t. This may work, but it is untested and is exactly the kind of hidden-assumption failure that would sink sim-to-real transfer, since real overheating or power-supply degradation is typically gradual and partial. The proposed concrete test—varying kτ across intermediate values—would settle whether the binary-conditioned modulation helps or hurts in the partial regime. Because this is an addressable experimental gap rather than a mathematical refutation, the conditional acceptance verdict remains appropriate, but the condition should explicitly require partial-fault evaluation and severity-aware ablations.","tokens_in":11191,"tokens_out":6854,"duration_ms":70263,"concrete_test":"Re-run the simulation evaluation of Table III for the weakened-motor condition with kτ ∈ {0.05, 0.15, 0.25} (and, as a boundary check, kτ = 0.0) while keeping all other settings identical; then report DreamFLEX vs DreamFLEX w/o modulation and FT-Net. If DreamFLEX's mean ATE is not below both baselines for every kτ, or if the w/o modulation ablation wins for intermediate kτ, the binary fault representation is insufficient for partial weakening and the headline claim should be narrowed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section II.B.2 defines a weakened motor by scaling torque with kτ sampled uniformly in [kL, 0.25] during training, but Section III.B fixes kτ = 0.0 for all 'Weakened Motor' entries in Table III. That is a completely disabled motor, not a weakened one, so the continuous partial-torque regime is never evaluated. The fault representation is a binary vector f_t ∈ {0,1}^n (Eq. 5), and the modulation layer (Eq. 9) conditions only on this vector, so it cannot distinguish 10% torque from 25% torque; severity must be inferred from observation history through the latent z_t. Moreover, the 'Faulty leg joint motion' reward (Table I) penalizes motion of any faulty joint, which is sensible for a locked/disabled joint but may encourage the policy to abandon a partially weakened leg that could still contribute. If real overheating or power degradation is gradual and partial, the binary-conditioned modulation may over-adapt, and the reported advantage over FT-Net/Kim et al. may not transfer.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DreamFLEX, a reinforcement-learning-based fault-tolerant quadruped locomotion controller built on the authors' DreamWaQ. DreamFLEX adds a failure estimation and modulation network (FEMNet) that, from a history of proprioceptive observations, jointly estimates body linear velocity, a binary joint fault vector, and a terrain latent vector; the fault vector then modulates the latent via an affine (FiLM-style) transformation before the policy consumes it. Training randomizes locked-joint (Eq. 3) and weakened-motor (Eq. 4) conditions with a failure curriculum (Eq. 6). In Isaac Gym, DreamFLEX is compared with Kim et al., FT-Net, and two ablations, reporting lower mean absolute tracking error (Table III); real-world video demonstrations on a Unitree Go1 in rough outdoor and indoor terrains are also presented.","tokens_in":11407,"tokens_out":7448,"duration_ms":68118,"significance":"If the quantitative claims survive scrutiny, the explicit fault-vector estimation plus modulation is a useful and interpretable design: it gives the controller an explicit indicator of where a failure has occurred and a learned mechanism for adapting the gait, and it is a natural extension of DreamWaQ. The failure curriculum is a practical contribution that could be reused. I found no circularity problem: the fault vector is supervised by ground-truth labels and the modulation parameters are learned by RL rather than fit to the evaluation metric. However, the current evaluation does not support all of the paper's claims: Table III has no variance information, the 'weakened motor' condition is never tested with partial torque loss, and the real-world evidence is qualitative. These are correctable with additional experiments, so the underlying idea merits revision rather than rejection.","major_comments":[{"comment":"Table III reports only point estimates of absolute tracking error, with no standard deviations, no number of seeds, and no significance tests, so the claim that DreamFLEX 'consistently outperforms' Kim et al. and FT-Net is not statistically supported. Because RL training and evaluation are stochastic, please run multiple seeds (e.g., at least five) and report mean ± std along with an appropriate significance test or effect size; the learning curves in Fig. 5 likewise need error bands before they can support the comparative claim.","section":"Table III; Fig. 5"},{"comment":"The 'Weakened Motor' entries in Table III are obtained with kτ = 0.0, which is complete torque loss, not a weakened motor in the sense of Section II.B.2 (kτ sampled from [kL, 0.25]). The real-world RR calf failure in Fig. 8 is described as a weakened motor but is also run with the fixed parameters of Section III.B, so the partial-torque regime is never demonstrated. Since f_t in Eq. (5) is binary and the modulation in Eq. (9) conditions only on f_t, there is also no evidence about how torque-loss severity would be represented; please evaluate a grid of kτ values (e.g., 0.05, 0.10, 0.20, 0.25) and report tracking error and fault-estimation accuracy for each. The 'Faulty leg joint motion' reward in Table I may additionally teach the policy to discard a partially functional leg, so this sweep is needed to verify the design in the partial-fault regime.","section":"Section III.B, Eq. (4), Table III"},{"comment":"The locked-joint evaluation fixes qcen = qdef, even though training samples qcen from a range that expands through the failure curriculum (Eq. 6). A single locked angle does not test the controller's ability to cope with joints locked at different positions, which is central to the locked-joint failure mode. Please report results for several qcen values, including positions away from the default stance.","section":"Section III.B, Eq. (6)"},{"comment":"The real-world experiments are presented as qualitative video stills and narrative descriptions only; no quantitative metrics are reported for the physical robot, such as number of trials, success rate, velocity tracking error, terrain-completion time, or fault-vector estimation accuracy, and no baselines are run on hardware. This supports a feasibility demonstration but not the abstract's claim that DreamFLEX outperforms existing methods in real-world scenarios; please add basic logged metrics or explicitly restrict the real-world claim to feasibility.","section":"Section III.C"}],"minor_comments":[{"comment":"The reward notation P(t^n_air,foot − t^n,des_air,foot) is not defined; specify whether P is a piecewise-linear operator, a projection, or a placeholder, and define v_xy,foot and the superscripts in the table.","section":"Section II.A.2, Table I"},{"comment":"In Eq. (7), the term 'LV AE' should be typeset as L_VAE; also define the VAE loss components explicitly rather than only by name.","section":"Section II.C.1"},{"comment":"Please specify the shapes of γ1 and γ2 and whether the affine transformation is applied elementwise to z_t; this matters for understanding how the modulation interacts with the latent dimension.","section":"Section II.C.2, Eq. (9)"},{"comment":"The ablation named 'DreamFLEX w/o joint fault vector' is said to be the same as DreamWaQ [10], but the name suggests only f_t is removed while the rest of FEMNet remains; clarify whether this ablation also removes the modulation layer and the fault-tolerant rewards.","section":"Section III.A"},{"comment":"Reference [24] lists the author as 'K. Kwamoto'; if the correct spelling is 'Kawamoto', please correct it.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript builds directly on the authors' own DreamWaQ and B-TMS; this is legitimate but makes the incremental nature worth monitoring. The main risk to the paper is not novelty but evaluation rigor: the missing partial-fault experiments and variance reporting are correctable, and I would encourage the editor to require them. I also note that no code or pretrained models are promised, which will limit reproducibility for a learning-based systems paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuine improvement over the DreamWaQ line, and the ablations make the contribution easy to see. The fault-vector estimation with BCE loss plus the FiLM-style modulation of the latent is a clean addition, and the paper earns credit for testing on real hardware, including an unplanned motor overheating event.\n\nWhat's new: FEMNet explicitly estimates a binary per-joint fault vector and uses it to modulate the latent terrain embedding before the policy. That differs from Kim et al.'s random joint masking and FT-Net's 1D CNN adaptor in a concrete way: the controller knows which joint is bad, and the modulation lets the policy change gait based on that knowledge. The ablations (w/o joint fault vector, w/o modulation) show each piece helps. That is the right way to present a learning-based contribution.\n\nWhere it slips: the paper never actually evaluates the 'weakened motor' claim. Training samples k_tau from [kL, 0.25], but evaluation fixes k_tau = 0.0, which is a fully disabled motor, not a weakened one (Eq. 4 gives zero torque). Since the fault vector is binary (Eq. 5), it cannot encode partial torque loss; the policy would have to infer severity from observation history via the latent. That may work, but Table III gives no evidence. A sweep over k_tau in {0.05, 0.1, 0.2} would close the gap. Also, Table III reports means only, no standard deviations or seeds, and the real-world section is qualitative video. These are fixable, but they matter because the headline claim is 'outperforms existing methods under weakened conditions.' The faulty-leg-joint-motion reward penalizes any motion of the faulty joint, which could push the policy to abandon a partially functional leg; worth a sentence discussing whether that is intended.\n\nThe stress-test note is right: the binary fault representation plus the fixed k_tau = 0.0 evaluation means partial torque failure is never demonstrated. But I would not call it a fatal flaw; the architecture could plausibly handle it via z_t, and training covers the range. It is an evaluation gap.\n\nWho this is for: anyone working on fault-tolerant legged locomotion or sim-to-real RL for quadrupeds. It deserves a serious referee and a conditional accept: ask for seeds/error bars, the partial-torque sweep, and one quantitative real-world measurement (e.g., success rate or tracking error over repeated runs).\n\nRecommendation: engage, but require the weakened-motor evaluation before signing off.","headline":"Solid sim-to-real fault-tolerant locomotion paper with a clear novelty, but the weakened-motor claim is trained, not actually evaluated.","tokens_in":11960,"tokens_out":2058,"would_cite":true,"duration_ms":21102,"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":"DreamFLEX is a learned controller that explicitly estimates which quadruped leg joints are locked or weakened, and uses that estimate to modulate the policy's internal state so the robot keeps walking on rough terrain where existing…","keywords":["quadrupedal robots","fault-tolerant locomotion","deep reinforcement learning","joint failure estimation","rough terrain traversal","proprioceptive state estimation","latent conditioning","locked joint and weakened motor"],"falsifier":"Inject a fault that matches neither model, such as a constant friction torque in a joint or intermittent motor dropout, then check whether the estimated fault vector stays accurate and velocity-tracking error stays below the no-modulation ablation; if estimation error rises sharply or the modulation fails to beat the ablation, the explicit-fault mechanism is not doing the claimed work.","tokens_in":11000,"feed_emoji":"🦿","tokens_out":12585,"duration_ms":110990,"temperature":0.7,"pith_summary":"DreamFLEX is a deep-reinforcement-learning controller that keeps a quadrupedal robot walking over rough terrain after a leg joint has failed. The paper's central claim is that explicitly estimating which joints are faulty, as a binary vector, and conditioning the policy's latent representation on that vector makes the controller more reliable than prior fault-tolerant locomotion approaches, both in simulation and on hardware. In the reported head-to-head simulation, the mean absolute velocity-tracking error under locked and weakened joint conditions is 0.1786 m/s, compared with 0.2848 m/s and 0.3710 m/s for the two comparison methods, and ablations that remove the fault vector or the modulation layer do worse. A sympathetic reading is that a robot can name its own hardware failure and change its gait accordingly, rather than stopping or falling when a joint locks mid-mission.","feed_headline":"Quadruped keeps walking over rough terrain after a leg joint fails","feed_subtitle":"A fault-detection network reshapes the gait in real time, clearing 10-15 cm stairs with a locked or overheated joint.","key_machinery":"The load-bearing component is the failure estimation and modulation network (FEMNet): an encoder-decoder that takes the five most recent proprioceptive observations and outputs the body linear velocity, a binary joint fault vector for all 12 joints, and a latent vector, the compressed representation of terrain and robot state. A decoder reconstructs the next observation so the latent space stays temporally consistent, with binary cross-entropy supervising fault detection and mean-squared-error supervising velocity. A modulation layer then maps the fault vector to affine parameters $\\gamma_1$ and $\\gamma_2$ and produces the conditioned latent $\\tilde{z} = \\gamma_1 \\cdot z + \\gamma_2$, which is what the policy network sees; this is the mechanism by which the controller changes its gait pattern depending on which joint has failed. The fault scenarios themselves are injected during training by clipping the commanded angle around a randomly sampled center, the locked joint case, or scaling the torque by a uniformly sampled factor, the weakened motor case, with a curriculum that widens the sampled range once task reward passes a threshold.","core_discovery":"The discovery is that a fault-tolerant locomotion policy improves when the fault is represented explicitly and used to reshape the latent code that the policy consumes, instead of being folded implicitly into the observation stream. During training the agent is randomly assigned a normal joint or a faulty joint, locked (commanded angle clipped to a small symmetrical range) or weakened (torque scaled by $k_{\\tau}$), and a failure curriculum gradually widens the severity. At deployment, the failure estimation and modulation network (FEMNet) reads the last five proprioceptive observations, estimates body velocity, a 12-dimensional binary joint fault vector, and a latent vector, then applies an affine modulation $\\tilde{z} = \\gamma_1 \\cdot z + \\gamma_2$ driven by the fault vector before the policy acts. The reported simulation benchmark places DreamFLEX's mean absolute velocity-tracking error at 0.1786 m/s under locked and weakened joints, below the 0.2848 and 0.3710 m/s of the compared approaches and below the two DreamFLEX ablations (0.2426 m/s without modulation, 0.2812 without the fault vector). Real-world trials on a quadruped platform confirm that the same controller climbs 10-15 cm steps and stairs and crosses gravel, roots, and slopes while a joint is locked or overheated, including an unplanned overheating failure.","pith_inferences":["Beyond the paper, the same explicit-fault-plus-conditioning recipe could apply to other degradation modes, such as sensor dropout, payload shift, or limb shortening, since the estimator only needs a labeled indicator and the policy only needs a conditioning signal.","Beyond the paper, the binary single-fault-per-agent training leaves open the case of two simultaneous faults; extending the fault vector to multi-label or severity-valued flags would be a direct stress test of the modulation mechanism.","Beyond the paper, isolating the fault-estimation benefit from the base terrain-imagination benefit would be a useful experiment: run the controller on precise foot-placement tasks where a wrong fault flag should appear as a visible misstep rather than as a small velocity error."],"forward_implications":["A mission does not have to stop when one actuator fails: the controller detects the fault from proprioception alone and rebalances onto the healthy legs, including on slopes, stairs, and uneven ground.","Because the fault vector is explicit, a single trained policy covers many fault locations and severities rather than needing a separate policy per broken joint; the paper shows performance for locked or weakened hip, thigh, and calf joints in front and rear legs.","The policy transfers to hardware without fine-tuning, and it handled a soft, deformable terrain that was absent from the training distribution, indicating out-of-distribution terrain generalization.","Explicit fault estimation plus latent modulation improves command tracking more than either removing the fault vector or feeding it directly to the policy, so the conditioning step is doing measurable work.","A real, unplanned overheating failure during summer testing was detected and handled, indicating the estimator is not only responding to the scripted fault injection."],"supporting_citations":[{"why":"Supplies the base proprioceptive locomotion architecture, observer, and latent terrain imagination that DreamFLEX extends.","marker":"[10]"},{"why":"Provides the implicit joint-state-estimation comparison baseline whose tracking errors are reported in Table III.","marker":"[26]"},{"why":"Provides the failure-recovery baseline using a 1D CNN adaptor that DreamFLEX is compared against.","marker":"[27]"},{"why":"Supplies the terrain curriculum and massively parallel simulation setup used to train all policies.","marker":"[31]"},{"why":"Supplies the concurrent-training idea of estimating explicit robot state from proprioception.","marker":"[34]"},{"why":"Defines the beta-VAE objective used to keep the latent vector a consistent implicit representation.","marker":"[35]"},{"why":"Inspires the affine conditioning of the latent vector by auxiliary information, adapted here to fault flags.","marker":"[37]"}],"fun_headline_variants":["Fault-aware controller climbs stairs with a locked joint","Robot reconfigures gait on the fly after joint lock or overheating","Explicit fault vector lets quadruped climb rough terrain mid-failure","Fault-aware modulation keeps robot stable on rough terrain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument stands on the assumption that the two simulated failure modes, a locked joint modeled as angle clipping and a weakened motor modeled as torque scaling, match the failures a quadruped actually suffers, so that the estimated binary fault vector is trustworthy in the field.","fun_headline_variants_meta":{"raw":{"variants":["Fault-aware controller climbs stairs with a locked joint","Robot reconfigures gait on the fly after joint lock or overheating","Explicit fault vector lets quadruped climb rough terrain mid-failure","Fault-aware modulation keeps robot stable on rough terrain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001114,"raw_usage":{"total_tokens":4670,"prompt_tokens":1004,"completion_tokens":3666,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":3597}},"tokens_in":620,"tokens_out":3666,"duration_ms":25359,"temperature":1.0,"reasoning_tokens":3597,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T17:49:38.223399+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inject a fault that matches neither model, such as a constant friction torque in a joint or intermittent motor dropout, then check whether the estimated fault vector stays accurate and velocity-tracking error stays below the no-modulation ablation; if estimation error rises sharply or the modulation fails to beat the ablation, the explicit-fault mechanism is not doing the claimed work.","supporting_citations":[{"cited_title":"DreamWaQ: Learning robust quadrupedal locomotion with implicit terrain imagination via deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the base proprioceptive locomotion architecture, observer, and latent terrain imagination that DreamFLEX extends."},{"cited_title":"Learning quadrupedal locomotion with impaired joints using random joint masking,","cited_arxiv_id":null,"evidence_quote":"Provides the implicit joint-state-estimation comparison baseline whose tracking errors are reported in Table III."},{"cited_title":"FT-Net: Learning failure recovery and fault-tolerant locomotion for quadruped robots,","cited_arxiv_id":null,"evidence_quote":"Provides the failure-recovery baseline using a 1D CNN adaptor that DreamFLEX is compared against."},{"cited_title":"Learning to walk in minutes using massively parallel deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the terrain curriculum and massively parallel simulation setup used to train all policies."},{"cited_title":"Concurrent training of a control policy and a state estimator for dynamic and robust legged locomotion,","cited_arxiv_id":null,"evidence_quote":"Supplies the concurrent-training idea of estimating explicit robot state from proprioception."},{"cited_title":"β-V AE: Learning basic visual concepts with a constrained variational framework,","cited_arxiv_id":null,"evidence_quote":"Defines the beta-VAE objective used to keep the latent vector a consistent implicit representation."},{"cited_title":"FiLM: Visual reasoning with a general conditioning layer,","cited_arxiv_id":null,"evidence_quote":"Inspires the affine conditioning of the latent vector by auxiliary information, adapted here to fault flags."}],"review_version":1}