acceptodds
Under review as a conference paper at ICLR 2027

What Makes Two Robot Failures the Same? Learning Recoverability Functions for Zero-Shot Robot Recovery

Abstract

Robotic recovery depends on a question that semantic failure labels do not directly answer: which corrective intervention will restore task progress from the current failed state? We introduce the recoverability function , which characterizes a natural failure trajectory by its response over a parameterized space of corrective interventions. RepairSpace learns an ordinal functional geometry in which nearby failures exhibit similar recoverability profiles, together with an intervention-conditioned predictor that estimates responses for unexecuted corrections. Recovery is performed by cost-constrained optimization over the predicted response surface. Across LIBERO-40 and RLBench-30, covering 70 tasks, 28,000 natural failures, three policy families, and 4.24M main-protocol recovery rollouts, RepairSpace reaches 0.824 success on unseen tasks and recovers 80.3% of failure-specific routing headroom. Under strictly matched supervision, the same 118M-parameter architecture improves unseen-task SR from 0.798 to 0.821 and cross-policy SR from 0.774 to 0.802. Its learned distance predicts held-out intervention transfer with correlation 0.842, compared with 0.694 for the matched full-trajectory response predictor. On held-out VLA-policy failures from unseen tasks, RepairSpace reaches 0.827 mean recovery success, compared with 0.784 for FLARE and 0.734 for REFLECT. The representation transfers across failure-generating policy families (0.806 success) and reaches 0.850 recovery success over eight real-robot failure configurations. Together, these results support recoverability as a functional coordinate system that connects failure representation to corrective action.

Primary area:
applications to robotics, autonomy, planning
Keywords:
robot learning, failure recovery, recoverability functions, representation learning, robotic manipulation, zero-shot generalization, intervention transfer

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_eb6UOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents REPAIRSPACE, a failure representation that captures robot failures based on their responses to interventions. The authors obtain recovery outcomes for a set of simulator checkpoints and learn both a failure representation and a response predictor, using an ordinal loss to bring failures with similar recovery responses closer together. During deployment, the approach uses a cost budget to choose a correction before running a fixed policy. The paper evaluates recovery for unseen tasks, failures caused by policies not used during training, held-out intervention parameters, and a small real-robot experiment.

Critical strengths
  1. The paper raises a good question: should two failures be considered similar if the same correction can fix both? Defining similarity based on repair results offers a concrete approach to categorize failures based on recovery, apart from perceptual or semantic category labels.
  2. The comparison with the response-only baseline is a key factor in this evaluation. When equal supervision, architecture, and search budget are used, the proposed objective increases unseen-task success from 79.8% to 82.1%. The ablations also quantify the benefit of confidence filtering and variance correction.
Critical weaknesses
  1. Implementation details are unclear, especially the architecture and training procedure of the continuation policy and how REFLECT and FLARE were adapted to the shared intervention library and budget. These details would be useful for reproducing and understanding the differences in the comparisons.
  2. The cross-policy experiments alter the policy that produces failures but hold the continuation policy and intervention library constant. Because recovery responses are reliant on both of these, these experiments demonstrate transfer within a specified recovery condition. While the authors note this as a limitation, it is not clear if the representation would remain useful with a changing continuation policy.
  3. The bound is about transferring an intervention from a nearby measured failure, whereas the deployed algorithm uses the learned response head and CEM search. The ordinal objective learns relative similarity over sampled interventions, whereas the bound assumes uniform control over response differences. The paper separates these settings, but the analysis doesn't lend much direct support to the deployed procedure.
  4. The real-robot evaluation tests eight planned failure configurations, each with ten trials, where REPAIRSPACE gets 68 successes while the response-only baseline gets 62. This is promising, but the small number of configurations and trials makes it hard to be sure how reliably this improvement would be observed. Providing an uncertainty estimate for the difference and testing more failures would be better.
Questions for authors
  1. What continuation policy is used after each correction, and how is it trained? Would the learned representation still transfer if we changed the continuation policy, rather than only the policy that generates the failures?
  2. Could the authors indicate the uncertainty in the real-robot improvement over the baseline, and whether trials were paired across methods under similar initial conditions? Demonstrating typical failures, attempted corrections, and resulting recoveries would aid interpretation.
Reviewer_1RojOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies how to represent robot failures so that the right recovery strategy can be selected after a policy fails. The authors argue that semantic labels of failure are not expressive enough to characterize which intervention would restore task progress, since failures with the same label may require different corrections and vice versa. To address this, they define a recoverability function that predicts the probability of task success conditioned on the failed state, a parameterized corrective intervention (such as retracting, regrasping, or locally realigning the gripper), and a fixed continuation policy that resumes the task after that intervention. They then learn an embedding in which failures with similar recoverability profiles are close, and select the correction for an unseen failure by optimizing the predicted success over cost-feasible interventions.

Critical strengths
  • The idea of expanding the representation of failure beyond the semantic descriptors and connecting it with the effect of corrective interventions seems an interesting approach.
  • The cost-matched headroom analysis (Table 13) provides an interesting insight. Choosing the correction per failure is worth about 0.3 SR over the best single intervention even among cheap local corrections, which motivates failure-conditioned recovery independent of this method.
Critical weaknesses
  • The main weakness is a lack of clarity that prevents assessing the results. The paper never describes how the cross-fitted oracle is computed, although it defines both ONR metrics and the headroom claim, and it does not explain how REFLECT and FLARE were reimplemented, trained, or mapped to the intervention library. This matters especially for FLARE, whose VLA fine-tuning and Reset skills appear excluded by the fixed continuation policy and B_main = 2.0. Several numbers are inconsistent (e.g., Table 19 versus Table 10), the figures are largely illustrative rather than informative, and the real-robot pipeline is not specified, so the comparisons behind the headline claims cannot be verified.

  • A second concern is computational cost. Training requires exact simulator snapshots and about 4.24M rollouts, and the learned geometry must be re-measured whenever the continuation policy or intervention library changes. This limits practicality outside simulation and makes the comparison with baselines that receive no comparable supervision difficult to interpret.

Questions for authors
  • How is the cross-fitted oracle computed? Over which candidate set, and with how many rollouts per test failure? Is it restricted to sampled candidates while RepairSpace searches continuously with CEM?
  • How were REFLECT and FLARE reimplemented on LIBERO-40 and RLBench-30? What training data did each receive, and how were their outputs mapped to the eight intervention families and their parameters? Was FLARE’s Reset mechanism available? Results with matched training data, or with FLARE at a budget that admits Reset, would materially affect my assessment.
  • Are the rollouts used to evaluate CEM-selected interventions (all methods, five seeds) included in the 4.24M total?
  • Is the functional geometry panel in Fig. 1 computed from real embeddings or purely illustrative? Fig. 2 is not referenced in the text and is hard to read; what should the reader take from it beyond Sec. 3?
  • The intervention families and parameter ranges (Table 4) are hand-designed. How were they chosen? How sensitive are results to the ranges, given that Table 30 only removes whole families? Does performance depend on the library matching the benchmark’s failure modes?
  • Success in contact-rich manipulation is often discontinuous in intervention parameters (e.g., grasp offset). What justifies the L_u-Lipschitz assumption, and how are L̂_z and L̂_u estimated in Table 28? The fit in Fig. 5d reports R² = 1.000 on five points.
  • Can the authors distinguish whether an intervention succeeds because it brings the continuation policy back in distribution, rather than because it addresses the specific failure? For example, they could compare against an intervention that returns to a nominal pre-failure state.
  • Training requires exact simulator snapshots and 48 to 768 rollouts per failure (5,940 GPU hours). How would response supervision be obtained outside simulation? For the real-robot study, was the model trained purely in simulation, and which failure detector and continuation policy were used?
Reviewer_h5ogOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work reframes failure similarity using recoverability. The proposed RepairSpace framework first uses a visual-temporal encoder and intervention-conditioned head to predict recovery outcomes. It then trains an ordinal objective with confidence-filtered triplets to group failures with similar recoverability patterns, forming ordinal representation geometry. At inference, CEM optimizes the predicted recovery surface for zero-shot repair. A regret bound accounts for coverage, representation mismatch and prediction error. Evaluated on 70 LIBERO and RLBench tasks with 28k natural failures and 4.24M rollouts, it achieves 82.4% unseen-task success, outperforming prior baselines, with strong correlation for intervention transfer and 85.0% success on real-robot setups.

Critical strengths
  1. The reformulation itself is valuable.
  2. Evaluation scale and fairness are top-tier for this area.
Critical weaknesses
  1. Thin core gain is obtained. Against the direct full-trajectory response predictor — same architecture, data and CEM budget, minus the ordinal geometry. Unseen-task SR improves only 2.6 points , real-robot 0.775→0.850 is better but reports no CI. The cost-benefit deserves franker discussion.
  2. The intervention space is eight hand-designed families and "zero-shot" happens entirely inside it. The open question is how the system degrades on repairs outside the library — gripper reorientation, asking a human. Coverage is addressed in theory, never verified empirically.
Questions for authors
  1. Please run leave-one-intervention-family-out: remove one family's response data at training and report how much that family's recovery SR drops at test. This directly probes the geometry's dependence on the library.
  2. How sensitive are results to failure-detection timing?
  3. How many trials per configuration for the real-robot 0.850, and why does the real column of Table 2 report no confidence intervals?
  4. What is the total compute for probing plus training, and how does deployment-time cost compare with FLARE/REFLECT?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.