acceptodds
Under review as a conference paper at ICLR 2027

EqC: Equilibrium Energy Control for Introspective Lifelong Robot Learning

11 views

Abstract

Long-term robot autonomy depends on more than effective action generation: a policy must recognize when its behavior is no longer sufficiently supported and use assistance to improve. We introduce **Equilibrium Energy Control (EqC)**, a framework for introspective lifelong robot learning built around an explicit conditional energy landscape. EqC casts generative action refinement as equilibrium-seeking optimization, while the geometry of the refinement process reveals whether the generated behavior remains consistent with the learned landscape. This competence signal enables Energy-Gated DAgger, where persistent energy-consistency violations trigger expert takeover and the resulting recovery trajectories are aggregated for subsequent policy updates. EqC integrates naturally into both Diffusion Transformer policies and flow-matching-based vision-language-action (VLA) policies. Across simulation and real-world bimanual manipulation, EqC matches or improves control performance and robustness while enabling timely intervention near the policy's competence boundary. A real-world study further shows how these interventions progressively expand the range of situations handled autonomously, providing a practical path from static generative controllers toward self-improving robot systems.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robotics, Embodied AI, Generative Control, VLA

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_85q6OpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a self-introspective approach in which a robot recognizes, during task execution, whether the current situation lies beyond its competence and requests assistance when it does. Unlike conventional flow-matching-based policies, action generation is reformulated as energy-based optimization. During training, the energy is modeled to increase along a prescribed pattern as the sample moves from noise toward the expert action; at deployment, the deviation from this pattern is used as an “uncertainty” signal. The method is applied to several VLAs and shown to improve control performance in both simulation and real-robot experiments. Furthermore, by using this signal to request human assistance and retrain on the collected data, the authors substantially improve real-robot success rates (96.7% and 98.3%), demonstrating the potential of metacognition-based robot learning.

Critical strengths
  1. Uncertainty signal obtained directly from the generation process: The main strength is that the “uncertainty” signal arises directly from the action-generation process, without requiring a separate monitoring model. The signal considers the entire generation trajectory rather than only the final output, and the experiments show that this yields more accurate detection (Table 14). This makes the approach particularly practical for large VLAs, where running multiple models is costly, and it suggests that the generation process itself can serve as a source of information about a model’s reliability.

  2. More effective intervention than human judgment: On the real robot, with the same amount of collected data, EqC-triggered intervention achieves much higher success than human-triggered intervention (96.7% vs. 58.3%). Since the two methods appear to share the same policy and differ only in how interventions are triggered (App. D.4), this is the most direct evidence that the signal identifies useful intervention points. A plausible explanation is that humans intervene only after a mistake becomes visible, whereas EqC can intervene earlier, when the model first begins to struggle, allowing it to collect data from states the model genuinely does not know how to handle.

Critical weaknesses
  1. Insufficient comparison between automated signals: In simulation, the paper does include a comparison in which only the signal differs on the same policy (Table 3b), but EqC’s advantage shrinks considerably in that setting (at m = 11, roughly two-thirds of the TPR gain is attributable to the policy difference). In the real-robot lifelong experiments, however, EqC uses the EqC-trained policy, while Diff-DAgger uses the original G0.5 policy (the initial success rates on Garment Folding also differ, 30% vs. 25%). It therefore remains unclear whether the real-robot improvement over Diff-DAgger is attributable to the signal.

  2. Unclear scope of failures the signal can detect: EqC treats a lack of smoothness in the action-generation process as a sign of uncertainty. As a result, it may miss “confidently wrong” cases, such as smoothly picking up the wrong block. Indeed, all failures detected by EqC in the paper (Figure 6), such as a block being pushed away or a sleeve not being grasped, involve gradual drift during execution. To claim that the robot is aware of its own competence, the paper should also show which failures the signal fails to detect.

Questions for authors
  1. [Same-policy comparison] In the real-robot lifelong experiments, what happens if the Diff-DAgger signal is applied to the EqC policy? Could the authors also state explicitly whether the human baseline uses the same policy as EqC? If the EqC signal still performs better under the same policy, this would further strengthen the paper’s central claim.

  2. [Detection timing] If EqC detects the onset of uncertainty earlier than humans do, can the authors show how much earlier the alarm is raised relative to the point at which the mistake becomes visible?

  3. [Missed failures] How does the EqC signal respond to “confidently wrong” failures, such as selecting the wrong object? What kinds of failures does the signal tend to miss? Depending on the answer, I will reassess the scope of the claim that the robot “knows its own competence.”

5 comments
AuthorsOpenReview ↗

Response to Reviewer 85q6

We thank the reviewer for the thoughtful and constructive feedback. We disentangle the issues below and address each with direct evidence and clarification.

Response to W1 Q1: Same-Policy Comparison

We thank the reviewer for this important suggestion. Our new same-controller closed-loop experiments demonstrate the effectiveness of EqC's intervention signal under matched lifelong-learning conditions.

We evaluate two additional automated gates, EqC + Diff-DAgger and EqC + Final-Step Gradient Gate, on real-world Garment Folding. All methods share the same initial EqC checkpoint, lifelong-learning protocol, and demonstration-count budget, differing in their intervention signals. Each method is evaluated over three independent runs, with 20 autonomous trials per checkpoint per run.

Table R1: Same-controller closed-loop lifelong learning on Garment Folding. Results report success rates (%) averaged over three independent runs.

MethodR0R1R2
EqC + Diff-DAgger30.048.381.7
EqC + Final-Step Gradient Gate30.043.361.7
EqC + Energy-Gated DAgger (ours)30.061.798.3

Energy-Gated DAgger consistently outperforms both alternative gates, achieving improvements of 16.6 and 36.6 percentage points at R2, respectively. These multi-run results provide direct empirical evidence that path-wise energy consistency contributes to more effective closed-loop lifelong learning, beyond improvements attributable to the initial policy architecture.

We also clarify that the human-intervention baseline uses the identical initial EqC policy and same EqC controller as our method, with only the intervention trigger changed. We will explicitly clarify this setting and include the additional same-controller comparisons in the revised manuscript.

Response to Q2: Detection Timing

We thank the reviewer for raising this question. Our real-world experiments demonstrate the learning benefits of EqC-triggered intervention.

Across three independent runs, EqC and human-triggered intervention collect an average of 13,883 vs. 13,141 expert-labeled frames on Block Stacking (five rounds), and 7,848 vs. 7,384 on Garment Folding (two rounds), respectively. With comparable amounts of expert supervision, EqC achieves substantially higher final success rates on both tasks: 96.7% vs 58.3% on Block Stacking and 98.3% vs 55.0% on Garment Folding. We do not claim that EqC intervenes earlier than humans, as the primary factor driving effective lifelong learning is the informativeness of the collected recovery demonstrations, rather than intervention earliness.

Intervention timing and learning value. For intervention-driven lifelong learning, the value of an intervention lies in the learning opportunities it creates. Recovery demonstrations from challenging states can provide informative supervision beyond expert trajectories in already mastered states. EqC enables the policy to assess its own competence directly from action generation, identifying states where additional expert supervision can meaningfully improve its behavior. This model-internal assessment allows intervention and data acquisition to be guided by the policy's learned competence. Human operators typically infer the policy's competence boundaries from observable behavior, whereas EqC's intrinsic competence assessment can better align expert supervision with the policy's actual learning needs.

Together, these results highlight the value of competence-aware intervention for acquiring informative recovery demonstrations and improving lifelong policy learning with comparable expert-data costs.

Response to W2/Q3: Missed Failures

Missed-failure examples. In Garment Folding, the robot may complete the folding motion while leaving one corner improperly folded, producing subtle visual differences that are difficult to detect from the robot's camera viewpoint. Nevertheless, such failure cases have only been observed occasionally in our real-world experiments. And these wrong behaviors reflect a common limitation of current VLA policies.

To mitigate these limitations, a promising direction is to complement EqC's model-internal signal with task-level verification, such as using a VLM to check final manipulation outcomes. Importantly, EqC provides proactive, path-wise competence assessment directly from action generation, without requiring an additional uncertainty estimator or waiting for observable task failures. Task-level verification can complement this capability by identifying semantic or outcome-level errors that internal consistency alone may miss. We view this integration as a promising future extension.

Finally, we thank you again for your time and consideration, and would appreciate it if you could consider raising the score in light of these clarifications and additional results.

Dear Reviewer,

Thank you for your insightful feedback, particularly regarding the independent contribution and limitations of EqC's competence-awareness signal. We have strengthened the revised manuscript in three directions:

Same-policy signal attribution. We conducted new real-world closed-loop comparisons using the same initial EqC controller and matched demonstration budgets. Across three independent runs, Energy-Gated DAgger achieves 98.3% final Garment Folding success, compared with 81.7% for EqC + Diff-DAgger and 61.7% for a final-step gradient gate. All methods start at 30% success, directly isolating the contribution of the intervention signal.

Intervention timing and learning value. We clarified that our central claim concerns the informativeness of competence-guided recovery data, rather than a measured earlier detection time relative to humans. With comparable expert-labeled frame counts, EqC achieves 96.7% vs. 58.3% success on Block Stacking and 98.3% vs. 55.0% on Garment Folding, compared with human-triggered intervention using the same controller.

More failure case study. We expanded the failure-case study to discuss semantic and outcome-level errors that internal energy consistency may miss, including incomplete garment folding despite internally consistent actions. We also discuss task-level verification using VLMs as a complementary direction.

We would sincerely appreciate it if you could revisit these new experiments and clarifications and let us know whether they address your key concerns before the discussion period concludes. Thank you again for your constructive suggestions.

Best regards,

The Authors

Reviewer_85q6OpenReview ↗

Response to the authors' rebuttal

Thank you for the additional experiments and the detailed clarifications. My comments on each question are below.

[Same-policy comparison] This concern appears to be resolved. In the new experiment, all methods started from the same EqC checkpoint (initial success rate of 30%) and differed only in the intervention signal. Under this setting, the final success rate on Garment Folding was 98.3% for Energy-Gated DAgger and 81.7% for the Diff-DAgger signal. The improvement can therefore be attributed to the signal itself rather than to the policy. The clarification that the human baseline also used the same EqC policy further supports this conclusion. However, this comparison is based on three runs of a single task. To assess the stability of the result, I would appreciate it if the authors could also report the variability across runs.

[Detection timing] In my review, I interpreted the advantage of EqC as coming from intervening "earlier" than humans. From the authors' response, I now understand that the key factor is not the timing of intervention but the informativeness of the collected recovery data. Given that EqC achieved substantially higher success than human-triggered intervention with a comparable amount of expert data (96.7% vs. 58.3% on Block Stacking and 98.3% vs. 55.0% on Garment Folding), this explanation seems reasonable. That said, the central idea of this paper is introspection, namely that the policy itself recognizes its own competence boundary. From this perspective, it would be very helpful to show at which stages of the task EqC-triggered interventions mainly occurred. For example, if the authors could show how these stages differ from those of human-triggered ones, and whether they coincide with the points where the policy actually struggles, the claim that the signal reflects the policy's competence boundary would become much clearer.

[Missed failures] The example of a garment with one corner not fully folded confirms that the signal can miss failures in which actions are generated smoothly but the outcome is wrong. I think this is a matter of the scope of the signal rather than a flaw in the method. Still, given the focus on introspection, it is important to delineate this scope clearly. For example, if the authors could report the fraction of failed real-robot trials in which the signal never exceeded the threshold, together with the types of these failures, it would become clear which failures the signal recognizes and which it misses. In addition, stating explicitly in the paper that the signal reflects inconsistency in the action-generation process rather than task success would convey the scope of the claim that the robot "knows its own competence" more accurately to readers.

Summary My main concern (same-policy comparison) has been resolved by the new experiment, and I regard the other two points as additional details that would further clarify the authors' claims. I will keep my current score of 3.

Reviewer_Yo9aOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Equilibrium Energy Control (EqC), which generates robot action chunks by gradient ascent on an explicit conditional energy function. It derives a path-wise energy-consistency score from the refinement trajectory and uses persistent violations to trigger expert intervention through Energy-Gated DAgger. The theoretical analysis characterizes stationary approximation, energy ascent, and consistency residuals, while explicitly acknowledging that internal consistency does not guarantee task correctness. Evaluation covers DiT and VLA policies, four simulation benchmarks, real-world bimanual manipulation, failure detection, and iterative intervention-driven learning.

Critical strengths
  • A coherent connection between control and introspection. EqC uses the same learned scalar landscape for action refinement and competence monitoring, deriving its detection signal from finite energy progress rather than adding a separately trained uncertainty model.

  • Substantial validation extending to physical learning loops. The evaluation spans multiple policy architectures, simulation benchmarks, and real-world manipulation, with matched multi-seed DiT reproductions and three independent intervention-learning runs.

Critical weaknesses
  • The closed-loop benefit of the proposed gate is not fully isolated. The comparison against G0.5 + Diff-DAgger changes both the controller and the detector; although an EqC + Diff-DAgger configuration is evaluated offline, it is absent from the physical learning comparison. A closed-loop comparison using the same EqC controller with an alternative automated gate, such as final-step gradient norm, would better establish how much improvement comes specifically from path-wise energy consistency.

  • The broader lifelong-learning claims exceed the current cross-task evidence. Appendix D.4 reinitializes the policy from a fixed checkpoint and retrains on the cumulative dataset each round, while Appendix D.5 studies forgetting recovery using one selected, best-performing run. These experiments support intervention-driven data aggregation, but do not yet establish a general retention advantage over ordinary cumulative replay across task sequences.

Questions for authors
  • Closed-loop attribution: Does the human-intervention baseline use the identical EqC controller and initialization, differing only in its trigger? Are same-controller closed-loop results available for alternative automated gates, and how do recovery duration and total expert-labeled frames compare across methods?

  • Hybrid detector implementation: How exactly is the Diff-DAgger score computed on the EqC-trained backbone? Since EqC learns a stationary energy gradient rather than the original time-conditioned flow field, please specify which field and target are compared and how the training-objective difference is accommodated.

  • Signal and deployment clarification: Equation (4) averages over (k=0,\ldots,K-1), whereas Appendix Eq. (62) uses (k=1,\ldots,K); which definition is implemented? Please also specify the physical persistence window, executed action-prefix length, and threshold calibration for (K=3), since these determine the actual intervention delay.

  • Cross-task evidence: Are the forgetting-recovery results available across the other initial training runs and against a matched cumulative-replay baseline? This would determine whether the cross-task finding supports a reproducible competence-monitoring advantage or should remain an illustrative case study.

4 comments
AuthorsOpenReview ↗

Response to Reviewer Yo9a (1/2)

We thank the reviewer for the thoughtful and constructive feedback. We disentangle the issues below and address each with direct evidence and clarification.

Response to W1 Q1: Closed-Loop Attribution

Human-intervention baseline. We clarify that the human-intervention baseline uses the identical EqC controller and initial checkpoint as our method, with only the intervention trigger changed.

Same-controller comparison. To isolate the contribution of the intervention signal, we conducted additional closed-loop experiments on real-world Garment Folding using EqC + Diff-DAgger and EqC + Final-Step Gradient Gate. All methods share the same initial EqC checkpoint, lifelong-learning protocol, and demonstration-count budget. Each method is evaluated over three independent runs, with 20 autonomous trials per checkpoint per run.

Table R1: Same-controller closed-loop lifelong learning on Garment Folding. Results report success rates (%) averaged over three independent runs.

MethodR0R1R2
EqC + Diff-DAgger30.048.381.7
EqC + Final-Step Gradient Gate30.043.361.7
EqC + Energy-Gated DAgger (ours)30.061.798.3

Energy-Gated DAgger consistently outperforms both alternative gates, achieving improvements of 16.6 and 36.6 percentage points at R2, respectively. These results demonstrate the contribution of path-wise energy consistency to closed-loop lifelong learning.

Expert-data costs. Across three runs, EqC, the original G0.5 + Diff-DAgger, and human intervention collect an average of 13,883, 12,128, and 13,141 expert-labeled frames per run on Block Stacking (five rounds), respectively. On Garment Folding (two rounds), the corresponding totals are 7,848, 6,946, and 7,384.

Together, these results highlight EqC's ability to identify competence-critical states through its intrinsic energy signal and acquire informative recovery demonstrations, leading to more effective intervention-driven lifelong learning. We will include these comparisons in the revised manuscript.

Response to Q2: Hybrid detector implementation

We thank the reviewer for raising this question. We adapt Diff-DAgger’s loss-based uncertainty estimator to EqC by evaluating EqC’s gradient-matching objective on the policy’s generated action chunk.

EqC defines the stationary energy and its gradient as

\qquad g_\theta(x,o)=\nabla_x E_\theta(x,o).$$ The scoring field is the energy gradient $g_\theta$, obtained by differentiating the scalar energy function. Given the generated action chunk $\hat a$, we treat it as a pseudo-label and independently sample $$t_k\sim\mathcal U(0,1), \qquad \epsilon_k\sim\mathcal N(0,I).$$ Following EqC’s training convention, we construct $$x_{t_k}=(1-t_k)\epsilon_k+t_k\hat a, \qquad u_k^*=c(t_k)(\hat a-\epsilon_k),$$ where $c(t)$ is the same rescaling function used during EqC training. The resulting score is $$s^{\mathrm{EqC}}(o)=\frac{1}{KHD_{\mathrm{act}}} \sum_{k=1}^{K} \left\| g_\theta(x_{t_k},o)-u_k^* \right\|^2,$$ where $H$ is the action horizon, $D_{\mathrm{act}}$ counts active action dimensions, and padded dimensions are masked out. The model timestep remains fixed at $t_{\mathrm{model}}=0$, while $t_k$ determines the interpolation and target rescaling. This adaptation preserves Diff-DAgger’s loss-based scoring principle using EqC’s stationary energy-gradient matching objective. Scoring occurs after action generation, leaving the generated action and model parameters unchanged. **This provides a same-controller baseline for isolating the contribution of EqC’s path-wise energy-consistency signal.**
AuthorsOpenReview ↗

Response to Reviewer Yo9a (2/2)

Response to Q3: Signal Definition and Physical Deployment

We thank the reviewer for carefully identifying this inconsistency and for highlighting the importance of the deployment parameters.

1. Energy-consistency definition. We apologize for the indexing inconsistency in Appendix Eq. (62). The correct definition is the one presented in Eq. (4), with the summation over k=0,…,K−1k=0,\ldots,K-1. Specifically, EqC performs KK gradient-ascent updates, producing a refinement trajectory x(0),…,x(K)x^{(0)},\ldots,x^{(K)}, where x(K)=x^x^{(K)}=\hat{x} serves as the final self-anchor. The energy-consistency score evaluates the KK pre-update states, including the initial noise sample x(0)x^{(0)}, relative to this final anchor. The terminal state x(K)x^{(K)} is excluded because it serves as the reference endpoint rather than an additional pre-update state. Moreover, its residual is zero when the normalization floor is inactive. Thus, for K=3K=3, the score averages the residuals at k=0,1,2k=0,1,2. We will correct Appendix Eq. (62) and the corresponding indexing in Appendix C.5 to ensure consistency with Eq. (4).

2. Physical deployment and intervention delay. We appreciate the reviewer's request for clarification. In physical deployment, EqC uses K=3K=3 gradient-ascent steps. The policy predicts a 32-step action chunk at a control frequency of 15 Hz and executes a prefix of 16 action steps before replanning. Expert takeover is triggered when the energy-consistency score exceeds the fixed threshold γ=0.2\gamma=0.2 for 3 consecutive replanning steps.

3. Threshold calibration. We calibrate on 40 real-world rollouts from Flower Insertion and Towel Folding, with 10 successful and 10 failed rollouts per task. This threshold provides useful separation between successful and failed executions and remains fixed across subsequent lifelong-learning experiments without task-specific retuning. We will clarify these deployment and calibration settings in the revised manuscript.

Response to W2 Q4: Cross-Task Evidence and Forgetting Recovery

We thank the reviewer for this important suggestion. Our new cross-task experiments demonstrate the benefits of EqC-guided recovery over ordinary cumulative replay across all three independent EqC training runs.

Experimental setting. Following Appendix F, we initialize each run from its Round 5 Block Stacking checkpoint (Table 17(a)). At R1, we collect six EqC-triggered Cup Stacking recovery demonstrations and update each policy using the cumulative dataset. Cup Stacking performance improves, while Block Stacking performance degrades across all three runs.

Matched cumulative-replay baseline. Both approaches start from the same R1 deployment policy and accumulated dataset and collect 12 additional demonstrations (six per task). Cumulative replay collects complete expert demonstrations from the beginning of each task, while EqC collects intervention-triggered recovery trajectories. This comparison isolates the data acquisition strategy under a matched demonstration-count budget.

Table R1: Cross-task adaptation and forgetting recovery across three independent runs. All entries report success rates (%) in the order Block Stacking / Cup Stacking.

R0R1R2: Cumulative ReplayR2: EqC-guided Recovery
Run 1100 / 060 / 5095 / 65100 / 100
Run 295 / 050 / 3570 / 6590 / 85
Run 395 / 070 / 2080 / 5095 / 70
Mean96.7 / 0.060.0 / 35.081.7 / 60.095.0 / 85.0

Results. EqC-guided recovery outperforms cumulative replay on both tasks across all three runs, achieving average improvements of 13.3 percentage points on Block Stacking and 25.0 percentage points on Cup Stacking. These results demonstrate the effectiveness of competence-aware recovery data acquisition for both new-task adaptation and recovery of previously acquired skills under matched demonstration-count budgets. We will include these additional experiments in the revised manuscript.

Finally, we thank you again for your time and consideration, and would appreciate it if you could consider raising the score and confidence in light of these clarifications and additional results.

Dear Reviewer,

Thank you for your constructive suggestions, which helped us strengthen the experimental attribution and reproducibility of EqC. We have addressed your weaknesses and questions through the following revisions:

Closed-loop attribution. We added three-run same-controller Garment Folding comparisons. Starting from identical EqC checkpoints and matched demonstration-count budgets, Energy-Gated DAgger achieves 98.3% final success, compared with 81.7% for EqC + Diff-DAgger and 61.7% for the final-step gradient gate. We also clarified that the human-intervention baseline uses the identical EqC controller, differing only in its intervention trigger.

Expert-data costs. We reported average expert-labeled frame counts for EqC, Diff-DAgger, and human intervention: 13,883 / 12,128 / 13,141 on Block Stacking and 7,848 / 6,946 / 7,384 on Garment Folding. This data indicates comparable expert-supervision costs across all three methods.

Implementation and deployment. We detailed the EqC-adapted Diff-DAgger scoring objective, corrected the refinement-indexing inconsistency, and specified physical deployment parameters, including K=3, a 16-step executed prefix, three consecutive threshold crossings, and calibration on 40 rollouts.

Cross-task evidence. Across all three runs, EqC-guided recovery outperforms matched cumulative replay, improving final Block Stacking and Cup Stacking success by 13.3 and 25.0 percentage points, respectively.

We would greatly appreciate your assessment of these additional results and clarifications, particularly whether they satisfactorily resolve your remaining questions before the discussion period ends. If these additional results and clarifications help address your remaining concerns, we sincerely hope they may also increase your confidence of our work. Thank you again for your valuable feedback.

Best regards,

The Authors

The authors have addressed most of my concerns, so we will mantain the current score of 4.

Reviewer_whBEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Equilibrium Energy Control (EqC), a framework for introspective lifelong robot learning. EqC learns an explicit conditional energy function to generate actions through iterative gradient-based refinement. It further introduces an energy-consistency score to estimate policy uncertainty from the refinement trajectory, which triggers expert intervention when potential failures are detected. The collected recovery demonstrations are incorporated into a DAgger-style lifelong learning process. Experiments on simulation benchmarks and real-world bimanual manipulation tasks evaluate action generation, failure detection, and lifelong policy improvement.

Critical strengths
  • The introduction is very well written and clearly motivates the problem setting, including how the paper measures action uncertainty, triggers expert intervention, and incorporates failure recovery data for lifelong learning.

  • EqC provides an intuitive uncertainty estimation mechanism by utilizing the energy information from the generative action refinement process itself, without requiring an additional uncertainty estimator. However, the novelty of this formulation remains somewhat questionable.

Critical weaknesses
  1. Limited novelty and insufficient empirical improvements

The novelty of EqC remains questionable, as the method builds heavily on existing equilibrium matching, energy-based optimization, and DAgger frameworks. Meanwhile, the improvements over strong baselines are relatively limited in several simulation benchmarks, such as LIBERO and VLABench, where the gains are often only a 3-8 percentage points. Although some experiments show larger improvements, the overall results do not consistently demonstrate a substantial advantage that clearly justifies the proposed formulation.

  1. Insufficient evidence that EqC provides an effective intervention signal

How reliably does the EqC score indicate that expert intervention would be beneficial? Although the paper reports failure-detection metrics, the connection between the uncertainty signal and the effectiveness of intervention deserves further clarification. The perturbation settings and evaluation composition discussed in Appendix B.4 and Table 8 should also be introduced more clearly in the main paper. There is no reason the consistency of denosing trajectory makes sense about how good the quality of the sample is. And the supervision signal c(t)c(t) is a defined schedule, the lack of explanation why a preset schedule can represent a gradient of energy is wondered.

Questions for authors
  • Eq. (1): How is the schedule c(t)c(t) justified as a target for the energy gradient? Since the true energy gradient is not directly available, what supports the effectiveness of this predefined schedule?

  • Eqs. (3)–(4): The energy-consistency formulation appears to measure deviation from a reference energy progression. How does this specifically characterize optimization geometry or equilibrium, rather than simply measuring consistency? A clearer connection to Figure 1 would be helpful.

  • Experimental settings: The perturbation variants and evaluation composition in Appendix B.4 and Table 8 are important for interpreting the failure-detection results. Could the authors clarify these settings in the main paper?

  • Inference latency: The reported per-replan latency is approximately 0.1 seconds, which is not negligible for robotic control. How does this fit into the actual control and replanning frequencies, particularly for systems operating at 30 Hz?

8 comments
AuthorsOpenReview ↗

Response to Reviewer whBE (1/3)

We thank the reviewer for the thoughtful and constructive feedback. We disentangle the issues below and address each with direct evidence and clarification.

Response to W1: Novelty and Empirical Improvements

EqC introduces a unified energy-based formulation for introspective lifelong robot learning, integrating action generation, intrinsic competence assessment, and targeted policy improvement through a single learned energy function.

Novelty

We acknowledge that EqC is inspired by Equilibrium Matching (EqM), which studies image generation and reports degraded generation quality with explicit energy formulations. However, we demonstrate that an explicit conditional energy function can serve as a competitive action generator for robotic control, matching or exceeding flow-matching baselines while enabling policy introspection. Specifically, we derive a reference energy-progression relationship from the gradient-matching objective and use energy-consistency residuals to assess competence along the action-refinement trajectory (Eqs. (3)–(4)). Energy-Gated DAgger translates this intrinsic signal into targeted expert intervention and recovery-data collection. Our experiments further demonstrate greater data efficiency than both human-intervention DAgger and Diff-DAgger, establishing the practical value of energy-based introspection for lifelong robot learning. EqC’s distinctive contribution is this functional coupling: the same learned energy function generates actions, assesses their reliability, and guides the acquisition of recovery experience for continual robot policy improvement.

Empirical Improvements

Improved results. EqC maintains competitive performance with flow-matching policies in simulation and achieves substantial gains in physical robotic manipulation. In simulation, EqC achieves competitive performance with flow-matching policies across all benchmarks, with notable improvements on VLABench (29.4% → 36.7%) and LIBERO-Plus (53.6% → 65.8%). Across three real-world bimanual tasks, EqC improves average success from 63.3% to 91.7%: Flower Insertion (45% → 80%), Towel Folding (80% → 95%), and Debris Sweeping (65% → 100%) (Table 3a). Furthermore, Energy-Gated DAgger achieves 96.7% and 98.3% success on Block Stacking and Garment Folding, respectively (Figure 4), demonstrating the effectiveness of competence-aware intervention in lifelong policy improvement.

Framework extensibility. Our new proof-of-concept experiment further demonstrates EqC's potential for automated lifelong learning. A prepared LLM-powered robot agent serves as the recovery expert, taking over when EqC triggers intervention. Incorporating six agent-generated recovery demonstrations improves autonomous Garment Folding success from 30% to 100% after one fine-tuning round. This demonstrates the feasibility of an automated intervention-and-learning loop. We will include this experiment in a new appendix of the revised manuscript.

Response to W2: Intervention Reliability, Evaluation Settings, and Energy Consistency

1. Reliability and effectiveness of the intervention signal. EqC's energy-consistency score measures deviations from a reference energy progression derived from expert-supervised gradient matching. A large deviation suggests that the generated action is not well supported by the learned energy landscape, indicating potentially unreliable behavior and motivating expert intervention. In our new same-controller closed-loop experiments on real-world Garment Folding, all methods start from the same initial EqC policy and follow matched demonstration budgets. After two learning rounds, Energy-Gated DAgger achieves 98.3% success, compared with 81.7% for EqC + Diff-DAgger and 61.7% for EqC + Final-Step Gradient Gate, averaged over three independent runs. We will introduce this experiement in detail in the revised manuscript.

2. LIBERO-Plus perturbations and evaluation composition.

LIBERO-Plus introduces controlled perturbations to the original LIBERO tasks. We address these experimental details in our response to Q3.

3. Energy consistency and supervision schedule.

EqC constructs the gradient supervision target c(t)(a−ϵ)c(t)(a-\epsilon) from expert demonstrations, without requiring a ground-truth energy gradient. The schedule c(t)=min⁡[4,20(1−t)]c(t)=\min[4,20(1-t)] bounds the supervision magnitude and decays to c(1)=0c(1)=0, encouraging the learned gradient to vanish near expert actions (Eq. (2)). Integrating this target along the demonstration bridge yields the reference energy progression (Eq. (3)). At inference, Eq. (4) measures deviations from this reference using the final refined action as a self-anchor, enabling intrinsic competence assessment.

AuthorsOpenReview ↗

Respond to Reviewer whBE (2/3)

Response to Q1: Justification of the Predefined Schedule c(t)c(t)

Following the equilibrium-gradient construction in EqM (https://arxiv.org/abs/2510.02300), EqC does not attempt to estimate an unknown ground-truth energy gradient. Instead, we construct a learnable energy landscape by prescribing the desired refinement behavior using expert demonstrations.

Specifically, EqC constructs an explicit scalar energy through the dot product Eθ(x,z)=⟨x,fθ(x,z)⟩E_\theta(x,z)=\langle x,f_\theta(x,z)\rangle, and supervises its gradient ∇xEθ(xt,z)\nabla_x E_\theta(x_t,z) to match c(t)(a−ϵ)c(t)(a-\epsilon), where xt=ϵ+t(a−ϵ)x_t=\epsilon+t(a-\epsilon). Here, expert demonstrations determine the desired refinement direction, while the predefined schedule c(t)=min⁡[4,20(1−t)]c(t)=\min[4,20(1-t)] controls its magnitude. The schedule remains bounded and decays to c(1)=0c(1)=0 near the expert endpoint, encouraging the learned energy gradient to vanish and supporting equilibrium-seeking refinement.

Thus, c(t)c(t) serves as an interpretable inductive bias for shaping the desired energy geometry, rather than an assumed ground-truth energy gradient. Unlike EqM's implicit field, EqC learns an explicit scalar energy for gradient-based action optimization and path-wise competence assessment, with monotonic energy ascent under suitable step-size conditions (Appendix C.4).

Response to Q2: Energy Consistency, Optimization Geometry, and Figure 1

We agree that Eqs. (3)–(4) measure deviations from a reference energy progression. However, this reference is derived from the gradient-matching objective used for action generation.

Specifically, integrating the prescribed gradient target along the demonstration bridge yields the reference energy-to-go relation in Eq. (3). At inference, since the expert action is unavailable, Eq. (4) uses the final refined action as a self-anchor and compares the realized normalized energy progression with this reference.

Importantly, the residual has a geometric interpretation beyond scalar consistency. As established in Appendix C.5 (Theorem 2), it captures accumulated gradient-field mismatch along the self-anchored chord and off-chord energy discrepancy, together with a normalization correction when applicable. Thus, the score reflects discrepancies in both the learned energy field and the refinement-path geometry.

Figure 1 illustrates this relationship: the middle panel visualizes deviations between the realized energy progression and the reference curve. The shaded region schematically represents these deviations, which are summarized by the mean absolute residual UΔEU_{\Delta E}. This enables action generation and path-wise competence assessment using the same learned energy function, without an additional uncertainty estimator.

AuthorsOpenReview ↗

Response to Reviewer whBE (3/3)

Response to Q3: LIBERO-Plus Experimental Settings and Evaluation Composition

We agree that the relationship between LIBERO and LIBERO-Plus, as well as the evaluation composition, should be made clearer in the main paper.

LIBERO-Plus is an established robustness benchmark that extends the original LIBERO tasks through controlled perturbations. We follow the official perturbation settings defined in LIBERO-Plus (https://arxiv.org/abs/2510.13626), which preserves the underlying task semantics while introducing seven controlled perturbation families: camera viewpoint, robot initialization, language, illumination, background, sensor noise, and object layout. As described in Appendix A.5, we evaluate the LIBERO-trained policy checkpoints on LIBERO and LIBERO-Plus.

For failure detection (Table 3b), we pool rollouts from eight evaluation conditions: the original LIBERO and the seven perturbation families of LIBERO in LIBERO-Plus. As detailed in Appendix B.4 and Table 8, we stratify the rollout pool by evaluation condition and use a disjoint 10% calibration / 90% evaluation split. The calibration subset contains 1,203 rollouts and provides approximately uniform coverage of the four LIBERO task suites. And it is used only to select thresholds for each detection window, which are then frozen for held-out evaluation. All compared detectors use the same split membership. We will explicitly clarify the relationship between LIBERO and LIBERO-Plus in our main text.

Response to Q4: latency.

The per-replan inference latency of EqC-adapted π0\pi_0 is approximately 0.1 s on LIBERO in simulation (Table 13). In real-world deployment, the standard flow-matching G0.5 policy has a latency of approximately 400 ms, while EqC-adapted G0.5 requires approximately 500 ms (Table 3a). This additional inference latency can be effectively addressed by Real-Time Chunking RTC (https://arxiv.org/abs/2512.05964). EqC supports asynchronous real-time control through RTC, which is a common and widely adopted approach in VLA deployment, and it enables asynchronous generation of the next action chunk while the current chunk is being executed, thereby overlapping model inference with robot execution. This effectively hides the additional inference latency of EqC during continuous robot execution.

RTC integration. During training, we sample delays of 0–12 steps within each 32-step action chunk, condition EqC on a clean action prefix, and apply the energy-gradient loss to the remaining suffix. During inference, EqC asynchronously generates the next action chunk while executing the current one, enabling continuous robot motion. For real-world validation, We integrate RTC into both G0.5 and EqC and evaluate them on the Galaxea A1Z dual-arm system operating at 15 Hz.

Table R1: Average task completion time on Towel Folding task in real world. Lower is better.

MethodWithout RTC (s)With RTC (s)
Galaxea G0.557.8838.68
Galaxea G0.5+EqC (ours)60.3437.28

RTC substantially reduces task completion time for both methods. EqC + RTC achieves 37.28 s completion time and maintains a 95% success rate (19/20 trials), with an average per-replan latency of 525 ms. These results demonstrate EqC's compatibility with efficient asynchronous control while preserving task performance.

Higher-frequency control. The same asynchronous approach supports higher control frequencies, such as 30 Hz, provided that the remaining action execution window covers the replanning latency with sufficient timing margin. This decouples the robot's control frequency from the model's inference frequency.

Overall, these results demonstrate EqC's practical compatibility with real-time robotic deployment, combining efficient asynchronous execution with strong manipulation performance.

Finally, we thank you again for your time and consideration, and would appreciate it if you could consider raising the score in light of these clarifications and additional results

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response

Summary of Revisions and Additional Evidence

We sincerely thank all reviewers for their constructive feedback. In response, we have strengthened the theoretical explanations, clarified the experimental protocols, and incorporated additional real-world evaluations into the revised manuscript. The main updates are summarized below.

Methodological clarification and theoretical consistency. We clarified the construction of EqC's explicit energy landscape, the role of the gradient supervision schedule, and the geometric interpretation of path-wise energy consistency. We also corrected the refinement-indexing inconsistency and detailed the hybrid Diff-DAgger detector.

Same-controller intervention comparisons. We conducted additional closed-loop Garment Folding experiments across three independent runs, using identical initial EqC controllers and matched demonstration budgets. Energy-Gated DAgger achieves 98.3% final success, compared with 81.7% for EqC + Diff-DAgger and 61.7% for the final-step gradient gate. We also reported expert-labeled frame counts.

Evaluation and deployment details. We clarified the LIBERO-Plus perturbation composition, calibration protocol, and physical intervention settings. We further integrated Real-Time Chunking (RTC), achieving 95% success and reducing Towel Folding completion time from 60.34 seconds to 37.28 seconds.

Cross-task lifelong learning. We extended the evaluation across three independent runs with a matched cumulative-replay baseline. EqC-guided recovery improves final Block Stacking and Cup Stacking success by 13.3 and 25.0 percentage points, respectively.

Framework extensibility and limitations. We demonstrated agent-assisted recovery learning, improving Garment Folding success from 30% to 100% in a proof-of-concept experiment, and added failure-case analyses discussing semantic and outcome-level errors that internal energy consistency may miss.

Together, these revisions provide clearer methodological grounding, stronger evidence for the contribution of EqC's intervention signal, and a more comprehensive assessment of its practical capabilities and limitations. We appreciate the reviewers' time and consideration of the revised work.

AuthorsOpenReview ↗

General Response to the Reviewers, Area Chairs, Senior Area Chairs, and Program Chairs

Dear Reviewers, Area Chairs, Senior Area Chairs, and Program Chairs,

We sincerely thank you for your constructive feedback. We have strengthened EqC with controlled experiments, methodological clarifications, and expanded real-world evaluations. All changes are included in the revised manuscript and highlighted in blue. The key updates are:

1. Agent-powered lifelong learning.

Most notably, we find our agent-powered lifelong learning experiment particularly promising. A prepared Astra robot agent provided six targeted recovery demonstrations, after which a single EqC fine-tuning round improved autonomous Garment Folding success from 30% to 100% (20/20 trials, without agent assistance during evaluation). We believe this proof of concept is especially meaningful, as it demonstrates the potential to close the loop between autonomous failure detection, agent-assisted recovery, and continual policy improvement.

2. Same-controller intervention comparisons.

Addressing Reviewers Yo9a and 85q6's concerns, we performed three independent Garment Folding runs with identical EqC controllers and matched demonstration budgets. Energy-Gated DAgger reached 98.3% final success, compared with 81.7% using Diff-DAgger gating and 61.7% using final-step gradient gating. We are pleased that these additional experiments addressed Reviewer 85q6's main concern and most of Reviewer Yo9a's concerns.

3. Cross-task lifelong learning.

Across three independent runs, EqC-guided recovery outperformed matched cumulative replay by 13.3 and 25.0 percentage points on Block Stacking and Cup Stacking, respectively, supporting competence-guided data acquisition across tasks.

4. Real-time deployment.

Addressing Reviewer whBE's latency concern, we integrated Real-Time Chunking (RTC). On physical Towel Folding, RTC reduced EqC completion time from 60.34 s to 37.28 s while maintaining 95% success.

5. Methodological and protocol clarifications.

To address Reviewer whBE's concerns about the training schedule and methodological novelty, we clarified explicit energy construction, training-schedule choices (supported by a new LIBERO-Goal ablation), and path-wise consistency geometry; We also distinguished EqC from Feng et al. (On the Guidance of Flow Matching, ICML 2025) and clarified LIBERO/LIBERO-Plus evaluation, threshold calibration, hybrid Diff-DAgger, and physical intervention protocols. Furthermore, in response to reviewer Yo9a's feedback, we corrected the indexing inconsistency in Appendix C.5 and clarified Eq. (4).

6. Intervention analysis and limitations.

Following Reviewer 85q6's suggestions, we compared human and EqC judgments on identical trajectories and analyzed physical failures. EqC detected 42 of 44 failed trials (95.5%), with missed cases clarifying the limits of internal energy-consistency assessment.

Together, these results reinforce EqC as a unified framework for robotic control, introspection, and intervention-driven lifelong learning. Thank you for your time and consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.