acceptodds
Under review as a conference paper at ICLR 2027

HAFT: History-Aware Flow Policy for Torque-Degraded Quadrupedal Locomotion

Abstract

Legged locomotion policies are typically optimized under nominal actuation conditions, implicitly assuming a stable relationship between control commands and the resulting robot dynamics. Actuator degradation violates this assumption by reducing joint-level actuator effectiveness, thereby altering the closed-loop dynamics and compromising coordinated whole-body locomotion. A fundamental challenge is that actuator effectiveness is not directly observable from instantaneous proprioception and must instead be inferred from the robot's temporal responses to its recent actions. Viewing this challenge through the lens of online system identification, we exploit recent observation–action history as informative evidence of the underlying actuation condition. Based on this insight, we propose HAFT, a history-conditioned framework for degradation-adaptive legged locomotion. HAFT infers latent actuator effectiveness from recent interactions and selectively aggregates temporal context according to the current robot state and inferred degradation condition. The resulting context is used to condition a flow-matching policy for joint action generation, enabling the policy to adapt whole-body coordination to changes in available actuation. Extensive experiments on quadrupedal locomotion under diverse actuator degradation patterns and severities demonstrate that HAFT consistently improves locomotion robustness and adaptation. Our code is available at https://anonymous.4open.science/r/HAFT-1B35.

Primary area:
reinforcement learning
Keywords:
reinforcement learning, quadruped robots, fault tolerance

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_EgddOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies quadrupedal velocity tracking on a Unitree Go2 when joint torques degrade. The degradation is modelled as a per-joint multiplicative scaling of nominal torque, τ = (1 − d)·τ̂, with d ∈ [0, 1] fixed for the rest of each episode after a single onset. HAFT encodes the last H = 10 completed transitions (observation, action, next observation, relative time offset) into tokens. Mean-pooled tokens are mapped to a 12-dimensional degradation estimate, supervised by the degradation that was active during the most recent completed transition. The current observation and the estimated degradation form a query for multi-head cross-attention over the transition tokens. The query is added back through a residual connection to give a context vector, which conditions a flow-matching policy that outputs 12 joint-position commands via a 64-step Euler integration. Training combines the flow-policy objective (ASPO, with a surrogate ratio built from flow-matching errors) with the degradation loss, and uses a curriculum that moves from ground-truth to estimated degradation. Experiments in Isaac Lab cover all 66 dual-joint pairs at two severities, 60 single-joint conditions, two velocity-tracking case studies, a contact-timing analysis, and ablations of the aggregation and flow components together with a curriculum-start sensitivity study. Baselines are MoE, FPO and PPO.

Critical strengths

S1. A physically motivated combination of explicit degradation estimates and transition-level history. The paper argues that the actuator estimate and the observed consequences of past actions carry different control-relevant information, and that the estimate should influence how history is selected, not only be concatenated as an input (Sec. 1, pp. 1–2). This is a clear design idea, and the causal alignment of the estimator's target with the most recent completed transition is handled carefully (Eq. 11).

S2. Systematic coverage of impairment configurations. The evaluation covers all 66 unordered dual-joint pairs at d = 0.5 and d = 1.0, and all 12 single joints at five severities (Figs. 3 and 7). This is broader than most fault-tolerance locomotion papers. The contact-timing analysis (Fig. 5) is a useful behavioural illustration of how support is reallocated after the front-right leg fails.

Critical weaknesses

W1. The ablations do not isolate the central mechanism. In Eq. 6 the query f_q([o; d̃]) contains the degradation estimate. In Eq. 8 the final context is c = LN(q + W_e e), so the same query also enters the output directly. Under the equations as written, removing d̃ from the query therefore removes it both from the attention and from the context that the actor receives. The HAFT_O ablation (Fig. 6, left; caption at pp. 8–9, lines 470–474) is described as removing "degradation-aware conditioning from history aggregation while retaining the interaction history". The text does not say whether the actor still receives d̃ in that variant. Its performance drop therefore cannot be attributed to degradation-guided history selection. It could equally reflect the loss of a direct fault input to the policy. A fair test needs a variant that keeps d̃ as a direct actor input and removes it only from the attention, and a variant that conditions attention on the current observation alone and then concatenates d̃ to the actor. Without these, the claim that attention selection driven by the estimate is useful is not supported. The paper also does not show that the learned attention weights depend on d̃, for example by plotting attention entropy or per-transition weights under different degradation estimates.

W2. The estimator's accuracy and its role in control are not evaluated. The paper's main claim depends on the degradation estimate d̂ being accurate enough to condition the policy. The only evidence is the illustrative heatmap in Fig. 1, and the training loss (Eq. 11). There are no held-out MAE or RMSE values, no post-onset response delay, no false-positive rate on healthy joints, and no comparison between an oracle degradation input and the estimate at test time. The curriculum-start study (Fig. 6, right) varies training only and does not test whether the policy depends on accurate identification at deployment. I would like to see oracle versus estimated conditioning at test time, and controlled noise or bias injected into d̂. If the oracle and estimated results are close, the value of explicit estimation needs to be reconsidered.

W3. Statistical reliability is not established. The number of training seeds, the number of evaluation episodes per condition, and the meaning of any variability shown are not stated anywhere in the text or appendix. Figures 3, 6 and 7 report single values without uncertainty. The margins over the strongest baseline are small in several places. My own averages from the printed values of Fig. 3 (equal weight over conditions, not author-reported) are 36.3 for HAFT against 35.8 for MoE at d = 0.5 over dual-joint pairs, and about 37.2 against 36.9 for MoE over single-joint conditions. Whether these differences exceed seed-to-seed variation cannot be determined from the paper. At full dual-joint loss HAFT's averages are higher (about 22.5 against 20.5 for FPO and 18.0 for MoE), which is more encouraging, but again no uncertainty is given.

W4. Several low-reward conditions are not discussed. Under complete loss of FLT+RLT, FRT+RRT and RLT+RRT, HAFT's reward in Fig. 3 is 3.49, 2.21 and 2.50. For FLT+RLT, HAFT is above MoE (1.70) but well below FPO (8.35). For FRT+RRT, HAFT is below both MoE (7.75) and FPO (2.62). For RLT+RRT, HAFT is above MoE (2.20) and FPO (1.59), but all three methods are very low. The paper says HAFT "retains higher rewards over a broad range" (p. 7) but does not discuss these cases. Reward alone cannot distinguish falling, stopping, or slow but stable locomotion. The authors should report survival rate, velocity error and recovery time for these conditions. Figure 4 (right column) also shows PPO failing to track the commanded x and y velocities before the fault is injected at t = 3 s. The paper does not explain this, and it is not stated how the two case-study episodes were chosen. The healthy condition (d = 0) is not reported for any baseline in Figs. 3 or 7, so the cost of the method when no fault is present is unknown.

W5. The flow-policy claim is argued rather than shown. The paper states that a diagonal Gaussian imposes conditionally independent variation across action dimensions, while a flow policy "can model dependencies among joint commands and accommodate non-Gaussian joint action distributions" (pp. 2–3, lines 82–85; Sec. 4.4, lines 289–291). The first statement is a correct property of diagonal Gaussians. The second is an expressivity statement about flow policies, which the paper does not test for the trained policy. Note also that the mean of a diagonal Gaussian is produced by a network that sees the full input, so it can already output coordinated joint commands. The evidence offered for the flow component is HAFT_F in Fig. 6 (left), which replaces the flow with a diagonal Gaussian. That variant also changes the objective and the training details, and the paper does not say what objective HAFT_F uses (the ASPO ratio in Eq. 14 is defined for flow-matching errors). The gain therefore cannot be attributed to conditional dependence, non-Gaussianity, capacity or optimization. A capacity-matched full-covariance Gaussian, and an analysis of the covariance of sampled actions under a fixed context, would address this.

W6. Scope of the degradation model. The degradation model in Eq. 1 is multiplicative and static after a single onset. Real actuator degradation is often non-linear and state-dependent. Thermal derating and current limits cause saturation of large torques, which a scaling model does not represent. Back-EMF and voltage limits make the usable torque depend on joint speed. Gear wear and friction increase can be additive and velocity-dependent. Degradation may also develop gradually and depend on the operating history. The paper does not discuss whether the linear model is adequate for these effects. I do not think real-robot experiments are necessary for a simulation paper, but the conclusions should be stated for the modelled setting. Several simulated variants would substantially strengthen the claims: torque saturation, speed-dependent derating, and slowly varying degradation. The authors should also say what the estimator would output under these variants and whether it would still be used for conditioning.

W7. Reproducibility and computation. Table 1 and Table 2 (Appendix D.1) give the network sizes, the Adam learning rate of 1e-4, the discount factor of 0.99, 4096 environments, 1500 iterations and a rollout length of 24. Several elements are still missing: M, w_deg, the ASPO parameters, the reward terms and their weights, the observation composition, the action scaling and PD gains, the termination rules, the curriculum boundaries N_start and N_end, p_deg, the baseline configurations, and the control frequency. It is also not stated whether baselines receive the interaction history, which matters for interpreting the comparison. For the joint update of the encoder and policy, the paper does not say which context the behaviour-policy flow-matching error uses. It is stored with the rollout, or recomputed with θ_old from the stored history, and whether it is stop-gradient. Action generation requires 64 Euler steps (Eq. 22–24), and no latency, hardware or control-rate figures are reported. Without these, the method cannot be reproduced or assessed for real-time use. At the time of my review, the anonymous repository returned HTTP 403 (github_app_access_required) for individual file requests, so I could not check the implementation against the paper. This is a fixable access problem and I do not draw any conclusion from it about the code itself.

Questions for authors
  1. HAFT_O. Does HAFT_O still give the actor the estimated degradation vector? If it does not, is the drop in Fig. 6 due to the attention selection, the loss of the direct input, or both? Please report a variant that keeps the direct input and removes it only from the attention query.

  2. Baselines and history. Do MoE, FPO and PPO receive the interaction history, and if so in what form? Please report a control that uses the current observation, or mean-pooled history, with d̂ concatenated to the actor input.

  3. Estimator accuracy. Please report held-out MAE or RMSE per joint and per severity, the post-onset response delay, and the false-positive rate on healthy joints. How does performance change when the policy receives the oracle degradation at test time, and when controlled noise or bias is added to d̂?

  4. Attention. Does the attention distribution over transitions change with the degradation estimate? A plot of attention entropy or per-transition weights under different estimates would answer this.

  5. Flow and optimization. What objective does HAFT_F use, and what are the capacity and exploration settings of HAFT, HAFT_F and the baselines? Is the baseline called FPO in Fig. 7 the same as FPO++ in Appendix E.1? How does a capacity-matched full-covariance Gaussian perform, and what correlations do the sampled actions show under a fixed context?

  6. Statistics. How many training seeds were used for each method, how many evaluation episodes per condition, and what do the shaded or thin traces in Figs. 6 and 8 represent? Please report per-condition means with confidence intervals, and state whether the gap of HAFT over MoE at d = 0.5 exceeds seed-to-seed variation.

  7. Failure cases. What causes the low rewards under FLT+RLT, FRT+RRT and RLT+RRT? Please report survival rate and velocity error for these conditions. How were the two episodes in Fig. 4 chosen, and why does PPO fail to track before the fault?

  8. Healthy condition. How do HAFT and each baseline perform when d = 0 for all joints? Is there a cost to the method when no degradation is present?

  9. Degradation distribution. What is p_deg in full, including the joint-count distribution, the severity range and the onset-time range? Are the evaluated configurations included in the training distribution, and if so are any held out?

  10. Model scope. How would the estimator and policy behave under torque saturation, speed-dependent derating or slowly varying degradation? Please include at least one such variant in simulation or state the limitation explicitly.

  11. Inference cost. What is the action-generation latency for 64 Euler steps on a stated hardware platform and control frequency? How does performance change with fewer integration steps?

  12. Reproduction. Please make the anonymous repository accessible, and provide M, w_deg, the ASPO settings, N_start and N_end, the reward and observation definitions, and the baseline configurations.

10 comments

We sincerely thank the reviewer for the detailed assessment and insightful suggestions. We have carefully revised the manuscript and conducted additional experiments to strengthen the evaluation of HAFT and clarify its methodological contributions.

W1 & Q1 & Q2 & Q4. Clarification of the ablation design and further evaluation of degradation-conditioned history aggregation.

We have clarified the definition of the original HAFT_O ablation. Specifically, HAFT_O removes the degradation estimator, its supervision loss, and the use of estimated degradation for policy conditioning, while retaining observation-conditioned attention over historical transitions. Therefore, HAFT_O evaluates the overall contribution of explicit degradation estimation and conditioning, rather than the isolated effect of degradation-guided attention. We have revised the corresponding description to avoid ambiguity.

To further examine the contributions of the proposed components, we have introduced two additional ablations in the revised Fig. 6(a). HAFT_A replaces cross-attention with mean pooling while preserving the history encoder, degradation estimator, and flow policy. Importantly, the degradation-conditioned query is retained in the residual context, allowing the policy to use estimated degradation without attention-based history selection. HAFT_H removes the history-based degradation-aware representation and conditions the flow policy solely on the current observation. Together with HAFT_O and HAFT_F, these variants provide a more detailed evaluation of the proposed architecture. The results show that the full HAFT achieves the highest late-stage training reward among the evaluated variants, supporting the contributions of explicit degradation conditioning, historical information, and attention-based aggregation. Furthermore, the sensitivity analysis in Fig. 6(b) shows that removing the auxiliary degradation estimation loss leads to lower late-stage rewards, providing additional evidence for the benefit of explicit degradation supervision.

Regarding the distinction between degradation-guided attention and direct policy conditioning, the original HAFT uses a shared query for both attention aggregation and residual context formation, as defined in Eqs. (6)--(8). Consequently, removing degradation information from this query affects both its contribution to attention and its direct contribution to the policy context. We therefore do not attribute the performance reduction of HAFT_O solely to degradation-guided attention. Instead, HAFT_A examines whether attention-based aggregation provides additional benefits over mean pooling while preserving degradation information in the policy context. This comparison supports the contribution of adaptive history aggregation, although it does not independently isolate the effect of degradation information on attention weights.

We have also clarified the input configurations of the baseline methods. All methods are evaluated under the same robot model, observation and action spaces, degradation settings, and training budget. The revised experimental description makes the differences in history processing explicit, while the HAFT ablations provide further comparisons of individual components.

Finally, we have added visualizations of the learned attention weights in Appendix Fig. 9 under healthy actuation, partial degradation, and complete torque loss of the FRT joint. The weights are recorded during closed-loop evaluation using the model's online degradation estimates. The results reveal different temporal attention patterns across degradation conditions, with several heads assigning greater weight to recent transitions under severe degradation. These observations show that HAFT adjusts its attention over historical transitions under different actuator conditions, although the changes cannot be attributed exclusively to the degradation estimate because the observations and interaction histories also vary.

W2 & Q3. Evaluation of Degradation Estimation and Its Impact on Control Performance

We have conducted two additional evaluations to examine how degradation estimation errors affect the control performance of HAFT. Specifically, we compare online degradation estimates with ground-truth degradation information and evaluate the policy's sensitivity to random noise and systematic bias. The results are presented in Appendix Figs. 7 and 8.

In Appendix Fig. 7, we replace the online degradation estimates with ground-truth values during evaluation while keeping the trained policy unchanged. Compared with the results obtained using estimated degradation, the two settings achieve broadly comparable rewards across most single- and dual-joint degradation configurations, and ground-truth conditioning does not consistently yield higher rewards. These findings suggest that the learned estimates provide effective conditioning information for locomotion control under the evaluated conditions. Importantly, both settings rely on degradation information for policy conditioning; thus, their comparable performance does not diminish the contribution of explicit degradation estimation.

We further investigate the effect of estimation errors in Appendix Fig. 8. Under additive Gaussian noise, the mean episodic reward remains relatively stable at standard deviations of 0.05 and 0.10, with a moderate reduction at 0.20. Under systematic bias, the highest reward is observed near zero bias, while larger positive or negative deviations result in lower performance. The shaded regions represent 95% confidence intervals. These results indicate that HAFT exhibits tolerance to moderate estimation errors, while larger or persistent deviations can adversely affect locomotion performance. Together, these additional experiments provide further evidence of how degradation estimation quality influences closed-loop control.

W3 & Q6. Statistical Reliability and Experimental Results

We thank the reviewer for the comments on statistical reporting. We clarify that all methods were trained with five independent random seeds, and each trained policy was evaluated over 256 episodes per degradation condition. The heatmaps in Fig. 3 report the mean episodic rewards, obtained by averaging the evaluation results within each training seed and then across the five seeds. This presentation facilitates comparison across different joint configurations and degradation severities.

Regarding the comparison between HAFT and MoE, their mean rewards are relatively close under moderate degradation (d=0.5d=0.5), particularly for certain single- and dual-joint configurations. We therefore interpret these results as indicating comparable performance in such conditions, without overemphasizing small differences in the reported means. Under complete dual-joint torque loss (d=1.0d=1.0), HAFT achieves larger mean reward margins in several configurations, indicating a more pronounced performance advantage under severe degradation. These observations highlight the differences in average performance across degradation severities and affected joint configurations.

Regarding the dashed lines in Figs. 6 and 8 of the original submission (corresponding to Figs. 6 and 12, respectively, in the revised manuscript), these are standard background grid lines of the plot axes. They do not represent individual training runs or statistical uncertainty, such as standard deviations or confidence intervals.

Reviewer_51sFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a framework for quadruped locomotion under actuation (torque) degradation. It estimates per-joint degradation and uses it to guide attention with a history of observations. Finally it uses a flow-based policy for action generation. Simulation experiments show better velocity tracking and robustness under simulated single and dual joint degradations compared to PPO, FPO, and MoE baselines.

Critical strengths
  1. The degradation-conditioned history aggregation is an interesting architectural idea, allowing the policy to retain information beyond just the estimated actuator degradation.

  2. The experiments cover a range of single and dual joint degradations, and the ablation studies provide some evidence that the proposed components improve performance.

Critical weaknesses
  1. Limited realism and generalization of the actuator degradation model: The paper models actuator degradation as multiplicative torque scaling. Within each episode, the torque randomly drops but always to a constant value which is then kept the same during the episode. Real actuator faults can involve torque saturation, thermal effects, and time-varying losses. It is unclear whether the proposed method generalizes to these scenarios. Furthermore, the authors should clarify how actuator torque limits are enforced relative to the degradation operation, as a position-controlled policy may otherwise compensate for partial torque scaling simply by increasing commanded position errors. The evaluation is also restricted to flat-terrain velocity tracking under the same class of degradation scenarios considered during training. Testing gradual or intermittent failures, different terrain conditions, and other dynamics perturbations would strengthen the claim. Finally, some baselines also exhibit poor tracking even before fault injection in Figure 4, making it difficult to separate improvements in nominal locomotion from improvements in fault adaptation.

  2. Several previous works as cited in the paper, including DreamFLEX [1] and ADAPT [2], have already demonstrated actuator-failure estimation and adaptive quadrupedal locomotion, with hardware experiments. The main novelty here appears to be combining degradation-conditioned attention with flow-based action generation. However, the paper does not sufficiently establish why this combination is necessary or substantially improves upon existing, simpler approaches.

Questions for authors
  1. How does HAFT compare against a standard PPO policy with the same history encoder and supervised degradation estimator, but without attention or flow-based action generation?

  2. How are actuator torque limits enforced in the simulator? Does the method generalize to torque clipping, gradual degradation, or changes in degradation severity within an episode?

  3. How does the proposed approach compare conceptually and empirically with existing fault-aware locomotion methods such as DreamFLEX and ADAPT? Those methods work on real hardware as well along with terrain variations, so a detailed comparison with baselines seems necessary to justify the additional complexity cost of the proposed method.

2 comments

We sincerely thank the Reviewer 51sF for the insightful comments and constructive suggestions. We have carefully revised the manuscript and supplemented the experimental evaluation to address the concerns raised in the review.

W1 & Q2. We clarify the actuator torque constraints and provide additional evaluations under gradual degradation, intermittent failures, and changes in degradation severity.

Our study focuses on reduced actuator effectiveness, modeled as a proportional decrease in joint torque. Specifically, the nominal torque generated by the PD controller is first constrained by the actuator torque limit, after which the degradation factor is applied. Consequently, the maximum available torque decreases proportionally with actuator effectiveness. Although the policy may partially compensate for torque loss by adjusting joint-position commands, it cannot restore the nominal maximum torque capacity of a degraded actuator. We have clarified this implementation in the Problem Formulation section (Pages 3–4 of the revised manuscript).

To further evaluate adaptation beyond the original setting of abrupt degradation with constant severity, we have conducted additional experiments involving gradual degradation, changes in degradation severity, and intermittent degradation. As shown in Appendix Fig. 11, HAFT generally maintains velocity tracking in the evaluated single-joint and dual-joint cases, with transient deviations and tracking fluctuations following changes in actuator effectiveness. Importantly, these evaluations use the trained policy without retraining, and the policy continues to rely on online degradation estimates rather than ground-truth degradation information. These results provide further evidence that HAFT can accommodate changes in degradation severity within an episode, beyond the constant-severity setting used during training.

Our study focuses on proportional torque degradation; other actuator fault mechanisms, such as thermal effects and speed-dependent derating, remain outside the evaluated scope. Regarding Fig. 4, PPO exhibits tracking errors before fault injection in the illustrated episode. Since actuator degradation has not yet occurred, these errors reflect its nominal tracking behavior in that episode and should not be attributed to the subsequent fault. Accordingly, the observed tracking differences should not be interpreted solely as differences in post-fault adaptation.

W2 & Q1 & Q3. Comparison with Existing Fault-Aware Methods and Contributions of HAFT

We thank the reviewer for the insightful comments. Previous studies, including DreamFLEX and ADAPT, have demonstrated the effectiveness of fault-aware locomotion through actuator fault estimation and adaptive control. Building upon these advances, HAFT focuses on using estimated joint-wise degradation to guide the aggregation of recent interaction history and condition flow-based joint action generation. The main contribution lies in integrating explicit degradation information with history-based adaptation and expressive action generation, rather than introducing fault estimation itself.

Following the reviewer's suggestion, we have incorporated DreamFLEX as an additional baseline and evaluated it under the same robot model, torque-degradation conditions, training budget, and evaluation protocol. The updated results in Fig. 3 show that HAFT achieves higher mean episodic rewards than DreamFLEX across most evaluated configurations, with particularly consistent advantages under partial torque degradation. These additional results strengthen the empirical comparison with existing fault-aware locomotion methods and provide further evidence of HAFT's effectiveness under different actuator degradation conditions.

To further examine the contributions of the proposed architecture, we have expanded the ablation study in Fig. 6. HAFT_O removes the degradation estimation and conditioning branch, HAFT_A replaces attention-based history aggregation with mean pooling, and HAFT_H removes the history-based representation entirely. In addition, HAFT_F retains the same history encoder, supervised degradation estimator, and attention module as HAFT, while replacing the flow policy with a Gaussian policy optimized using PPO. The full HAFT achieves higher late-stage training rewards than these variants, supporting the benefits of degradation-aware history representation and flow-based action generation. We also report the inference cost of the flow policy in Appendix Table 3, with a measured mean actor inference latency of 1.066 ms using 64 Euler steps. Together, the additional baseline comparison, component ablations, and inference measurements provide further empirical support for the proposed design and its computational feasibility under the evaluated simulation setting.

Reviewer_9iVrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes HAFT, a history-aware flow policy for quadrupedal locomotion under joint torque degradation. The method uses recent observation, action, and next-observation transitions to estimate the degradation level of each joint. The estimated degradation and current observation are then used as query tokens to cross-attend over historical transitions, producing a context representation for a flow-matching policy that generates whole-body joint-position commands. During training, the degradation estimator is supervised using ground-truth actuator degradation, and a curriculum gradually replaces ground-truth degradation with predicted estimates. The flow policy is trained with an on-policy RL objective. The authors evaluate HAFT on a Unitree Go2 in Isaac Lab under randomized torque degradation, with comparisons against PPO, Flow Policy Optimization (FPO), and a mixture-of-experts (MoE) controller. HAFT generally achieves higher locomotion rewards, particularly under severe dual-joint failures.

Critical strengths
  1. Reasonable motivation and architecture.

    Combining explicit actuator degradation estimates with historical action-response information is a sensible approach to adaptation under partially observable actuator failures.

  2. Broad evaluation of actuator failures.

    The experiments cover many joint combinations and degradation severities, with promising results under severe dual-joint torque loss (Figures 3 and 7).

Critical weaknesses
  1. Ablation design

    The main novelty appears to be using estimated degradation to guide attention over historical transitions. However, the HAFT_O ablation is not very clear on exactly what it removes. I am not sure what exactly it removes. I think there might need to be several separate ablations of the components:

    • the degradation prediction loss
    • whether the degradation prediction is concatenated with the history obs to form the query.

    Comparisons with matched history-conditioned policies and direct degradation conditioning would better establish the contribution.

  2. Limited Baselines

    The paper primarily compares HAFT against PPO, FPO, and MoE, but the main contribution concerns degradation-aware history representations rather than policy optimization. More relevant methods, such as LocoFormer [1], DreamFLEX [2], and FT-Net [3], are already cited but not experimentally compared. LocoFormer demonstrates long-context adaptation to motor failures, while DreamFLEX explicitly estimates joint faults and uses them to condition locomotion. Comparisons with these methods under the same degradation settings and training budgets are important to establish the advantage of HAFT's proposed architecture.

  3. Limited analysis of degradation estimation and adaptation

    Although online degradation estimation is a central component, the paper does not quantitatively evaluate estimation accuracy, fault-detection latency, or adaptation/recovery time. The velocity-tracking plots only show selected examples. It is therefore unclear how reliably and quickly the system identifies degradation, or how much the estimation quality affects control performance.

  4. Insufficient implementation and experiment details

    The paper omits several important details, including the reward formulation, the number of cross-attention layers, and the history inputs and architectures of the baselines. The evaluation primarily reports episodic rewards without uncertainty estimates, evaluation seed counts, or aggregate velocity-tracking errors, fall rates, and recovery times. The flow policy inference latency or control frequency are also not reported, making it hard to check if this is viable for physical robot deployment.

[1] M. Liu, D. Pathak, and A. Agarwal, “LocoFormer: Generalist Locomotion via Long-context Adaptation,” in Proc. Conference on Robot Learning (CoRL)

[2] S. Lee et al., “DreamFLEX: Learning Fault-Aware Quadrupedal Locomotion Controller for Anomaly Situation in Rough Terrains,” in Proc. IEEE International Conference on Robotics and Automation (ICRA)

[3] Z. Luo, E. Xiao, and P. Lu, “FT-Net: Learning Failure Recovery and Fault-Tolerant Locomotion for Quadruped Robots,” IEEE Robotics and Automation Letters

Questions for authors
  1. Can the authors provide an ablation that preserves the same degradation estimate as a policy input while removing only degradation-conditioned attention?
4 comments

We sincerely thank the reviewer for the insightful and constructive feedback. We have carefully revised the manuscript and expanded the experimental evaluation to address the concerns raised in the review.

W1 & Q1. We have clarified the original ablation settings and introduced additional experiments to better distinguish the contributions of degradation estimation, historical information, and attention-based aggregation.

We have clarified the definition of the original HAFT_O ablation. Specifically, HAFT_O removes the degradation estimator, its supervision loss, and the use of estimated degradation for policy conditioning, while retaining observation-conditioned attention over historical transitions. Therefore, HAFT_O evaluates the overall contribution of explicit degradation estimation and conditioning, rather than the isolated effect of degradation-guided attention. We have revised the corresponding description to make this distinction explicit.

To further examine the contributions of the proposed components, we have introduced two additional ablation variants in the revised Fig. 6(a). HAFT_A replaces cross-attention with mean pooling while retaining the history encoder, degradation estimator, and flow policy. Importantly, the estimated degradation remains available to the policy through the residual context, allowing us to evaluate the contribution of attention-based history aggregation beyond direct degradation conditioning. HAFT_H removes the history encoder, degradation estimator, and attention module, conditioning the flow policy solely on the current observation. Together with the original HAFT_O and HAFT_F variants, these experiments provide a more comprehensive assessment of the proposed components. Furthermore, we have extended the sensitivity analysis in Fig. 6(b) to include the zero-weight setting, which removes the auxiliary degradation estimation loss while preserving the remaining architecture.

As shown in the revised Fig. 6(a), the full HAFT achieves the highest late-stage training reward among the evaluated variants. Removing the degradation estimation branch or replacing the flow policy with PPO results in lower rewards, while HAFT_A and HAFT_H exhibit more pronounced performance deterioration during later training. These findings support the contributions of explicit degradation conditioning and attention-based history aggregation to policy learning. The additional sensitivity results in Fig. 6(b) further show that removing the degradation estimation loss leads to lower late-stage rewards, supporting the benefit of explicit degradation supervision. We have accordingly revised the figure caption and experimental discussion to clarify the ablation settings and their corresponding findings.

W2. Limited Baselines: Additional Comparison with DreamFLEX

We thank the reviewer for this valuable suggestion. Following this recommendation, we have incorporated DreamFLEX as an additional baseline to strengthen the comparison with existing fault-aware locomotion methods. DreamFLEX explicitly estimates joint failures and uses the estimated fault information to adapt locomotion control, making it particularly relevant to the degradation-aware design of HAFT. We evaluate DreamFLEX under the same robot model, locomotion task, degradation settings, and evaluation protocol as the other methods.

As shown in the revised Fig. 3, HAFT achieves higher mean episodic rewards than DreamFLEX across most evaluated degradation configurations, with particularly consistent advantages under partial torque degradation. Under complete torque loss, HAFT also maintains favorable performance across a broad range of joint configurations. These additional results provide further empirical evidence of HAFT's effectiveness in adapting locomotion to different actuator degradation conditions and strengthen the experimental comparison with existing fault-aware approaches.

W3. Limited Analysis of Degradation Estimation and Adaptation

We have extended the experimental evaluation to further examine the contribution of online degradation estimation and its influence on locomotion performance, including additional analyses of degradation supervision, ground-truth conditioning, and robustness to estimation errors.

As shown in Fig. 6(b), removing the auxiliary degradation estimation loss results in lower rewards during the later stages of training, as the curriculum progressively shifts from ground-truth degradation to estimated values. Together with the HAFT_O ablation, this result supports the contribution of explicit degradation estimation and conditioning to policy learning.

To further assess how estimation quality affects control performance, we compare the use of online degradation estimates with ground-truth degradation information under single- and dual-joint degradation conditions. As shown in Appendix Fig. 8, both settings achieve broadly comparable rewards across most evaluated configurations, and access to ground-truth degradation does not consistently improve performance. This suggests that the learned estimates provide effective conditioning information for locomotion control under the evaluated conditions. We further examine the policy's sensitivity to estimation errors in Appendix Fig. 9. The results show that performance remains relatively stable under moderate random noise, whereas larger perturbations and systematic biases lead to reduced rewards. These findings provide additional evidence of the relationship between degradation estimation quality and closed-loop control performance.

Furthermore, the velocity-tracking results in Fig. 4, together with the additional evaluations under gradual and intermittent degradation, illustrate HAFT's ability to adapt its locomotion behavior to changes in actuator effectiveness. These experiments provide complementary evidence of online adaptation by showing the policy's response before and after degradation. Although these trajectories do not directly quantify fault-detection latency or recovery time, they offer insight into the closed-loop control behavior under different degradation conditions.

W4. Implementation Details, Experimental Evaluation, and Inference Efficiency

We thank the reviewer for the constructive comments. We have expanded the implementation details in the revised appendix and restored access to the anonymous code repository. Specifically, Appendix Table 1 provides additional architectural details, including the number of cross-attention layers (1). Appendix Table 2 further specifies the training configurations, including the number of flow-matching samples per action (M=16M=16), degradation supervision weight (wdeg=0.1w_{\mathrm{deg}}=0.1), ASPO parameter (ϵ=0.05\epsilon=0.05), and degradation-conditioning curriculum (iterations 600–1000). These additions complement the original implementation details and facilitate reproduction of the proposed method.

Regarding experimental evaluation, we clarify that the heatmaps in Fig. 3 report mean episodic rewards aggregated over repeated evaluation runs, rather than results from individual episodes. The heatmaps present the average performance across different degradation configurations to facilitate comparison among methods. We interpret the results as comparative performance trends and do not infer statistical significance from small differences in mean rewards alone.

We have also added a quantitative evaluation of flow-policy inference latency in Appendix Table 3. The measurements cover Euler integration steps of 4, 8, 16, 32, 64, and 128. At the default setting of 64 steps, the mean actor inference latency is 1.066 ms, with a standard deviation of 0.029 ms, corresponding to approximately 5.33% of a 20 ms reference control period. Furthermore, the integration-step sensitivity analysis in Fig. 6(c) shows comparable training performance across the evaluated settings from 16 to 128 steps. These results provide an empirical assessment of the inference cost and its relationship with policy learning, while the reported latency is limited to actor inference rather than the complete control pipeline.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.