- 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
-
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.
-
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.
-
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?
-
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.
-
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?
-
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.
-
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?
-
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?
-
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?
-
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.
-
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?
-
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
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 (), 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 (), 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.
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.