acceptodds
Under review as a conference paper at ICLR 2027

Learning to Recover: Iteratively Internalizing Agent Corrections into VLAs for Robotic Manipulation

Abstract

Vision-language-action (VLA) models acquire manipulation skills through large-scale training, while reasoning agents can use tools and policies to handle situations beyond those skills. Yet successful assistance does not itself improve the underlying policy, leaving subsequent executions dependent on renewed online reasoning. We present Learning to Recover (L2R), an agent-to-policy learning framework that iteratively internalizes agent corrections into reusable recovery capabilities. At natural failures, the agent constructs and verifies corrections; lightweight offline updates then teach the policy to reuse them without online agent assistance. Rather than imitating entire intervention records, L2R selectively learns valid corrective actions and when to invoke them through conditional low-rank adaptation of a frozen VLA and a lightweight local controller. Further rounds target failures of the updated policy, allowing recovery learning to follow the policy's changing behavior. On ten VLABench task families across five tracks, a single round trained only on in-distribution corrections achieves 52.0% conditional recovery, compared with 31.8% for standard LoRA. Three rounds improve policy-only task success from 12.6% to 27.4% on ten RoboDojo tasks and from 33.3% to 68.3% on three real-world tasks. These results show how limited agent corrections can be progressively internalized into autonomous recovery capabilities. Videos are available on our project page.

Primary area:
applications to robotics, autonomy, planning
Keywords:
robot learning, vision-language-action models, agentic recovery, parameter-efficient adaptation, continual learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_tTaAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes the L2R framework, a recovery pipeline that enables a VLA policy to self-recover from failures. During data collection, a tool-using agent takes over failed executions and curates verified recovery trajectories. By re-annotating these samples into which actions to imitate and when to trigger recovery, they train a lightweight adapter and controller on top of the frozen base policy, so the tuned policy can recover without calling the agent at deployment. Each round collects new failures from the updated policy, and experimental results show improvement after rounds of recovering.

Critical strengths
  1. Technically solid and soundness paper, evaluated on both simulations and real-robot.
Critical weaknesses
  1. Transferability. Each task learns its own adapter and controller, so recovery only works on a fixed task. The core recovery ability is not shown to transfer to other tasks. The authors should report where transfer stops. For example, on which variants of the target task does recovery still work?

  2. Weak baselines. Most baselines in the experiments are not self-recovery frameworks. To show that the framework adds real value, the authors should compare against baselines that use the same VLM agent, the same annotations, and the same recovery trials. For example: (a) a simple baseline that keeps only the base policy's own successful rollouts and fine-tunes on them, with no agent correction; (b) LoRA-BC trained with the same recovery conditioning and the same amount of data as each round (90/180/270 trajectories), not R1 data only; (c) existing self-recovery or self-improvement frameworks, such as HG-DAgger-style intervention learning, PLD, FailSafe, or RePO-VLA; (d) the tool-using agent controlling the robot directly online, with its success rate, latency, and cost reported as an upper bound. Without a complete and fair comparison, we cannot tell whether the gain comes from the framework or just from extra compute and the stronger abilities of the off-the-shelf tool-using VLM agent. The pipeline also relies on many separate CV tools for state estimation, which adds to this concern.

  3. Policy-agnostic experiments. The authors should test whether the framework works across different base policies and different reasoning agents. A clear relationship between the framework's effectiveness and the base model's capability is expected.

Questions for authors
  1. Controller robustness under distribution shift. The VLABench shifted tracks provide the recovery boundaries and routes, so they only test the adapter. The GRU controller is never evaluated under shift. Since it is trained per task on 90 recoveries plus nominal traces using mean-pooled visual features, I expect entry and route prediction to be the fragile part. Could you run the full policy (controller + adapter) from the start on the VLABench shifted tracks, or on the randomized RoboDojo episodes, and report entry precision/recall and route accuracy against agent labels, separately for ID and shifted conditions? I would also like to see whether the GRU controller transfers across conceptually similar tasks, for example from hanging headphones to hanging a towel or a mug.

  2. Sensitivity to controller design. The controller relies on several hand-set choices: an 8-query history, entry/exit thresholds, consecutive confirmations, and holding rules. None of them are ablated. How does from-start success change with history length (e.g., 1/4/8/16 queries) and with the entry threshold? A simple non-recurrent baseline, such as an MLP on the current query only, would also show whether the GRU's history matters.

  3. How are the route types defined? Is there a clear taxonomy of failure types, and can it be extended to new failure types?

  4. On RoboDojo, performance saturates after 2–3 rounds of recovery. What bounds the final performance? Is it bounded by the agent's direct control (an upper bound), or by the base policy's capability?

5 comments
AuthorsOpenReview ↗

Official Comment by Authors

We appreciate the time and care you put into this review. Your concrete suggestions helped us strengthen the evaluation; the additional baseline, transfer, model and controller studies are in Appendix E of the revised manuscript.

W1: Transferability—“where transfer stops.”

Response: We transferred frozen R1 controllers and adapters together, without target-task recovery training.

Table 1. Cross-task transfer of frozen R1 recovery modules.

SourceTargetBase SRTransferred SRGain (pp)Target-trained R1 SR
Stack BowlsStack Blocks10% (5/50)16% (8/50)+620% (10/50)
Stack BowlsBuild Tower0% (0/50)4% (2/50)+410% (5/50)
Insert TubesFill Pen Holder6% (3/50)12% (6/50)+616% (8/50)

Each target uses the same 50 initializations across configurations. The frozen modules improve success on all three targets by 4–6 percentage points, showing useful recovery reuse without target corrections. Target-trained R1 remains 4–6 points higher. Together with the 4% absolute success on Build Tower, this indicates that related-task transfer helps, while more complex assembly still benefits from target-specific recovery learning.

AuthorsOpenReview ↗

Official Comment by Authors

W2(a): Weak baselines—“the base policy's own successful rollouts,” with “no agent correction.”

Response: The baseline and model comparisons use Stack Bowls and Insert Tubes, with 50 from-start episodes per task and configuration. LoRA used 90 Base successes per task.

Table 2. Learning from Base successes only.

TaskTraining successesBase SRSuccess-only LoRA SRL2R-R1 SR
Bowls9062% (31/50)64% (32/50)76% (38/50)
Tubes908% (4/50)8% (4/50)20% (10/50)

Success-only learning adds one successful trial on Bowls and none on Tubes. L2R-R1 reaches 76% and 20%, exceeding this baseline by 12 percentage points on both tasks. In this comparison, the L2R recovery-learning setup improves task success more than success-only fine-tuning.

W2(b): Weak baselines—“the same recovery conditioning” and “90/180/270 trajectories.”

Response: Appendix D compares condition-aware LoRA and L2R at R1 on four tasks, with macro success rates of 25.0% and 35.0%. We extended the comparison to 90/180/270 recovery trajectories on Bowls and Tubes.

Both methods use the same corrections, action targets, masks and replay, with matched adapter-update budgets. The comparison LoRA receives the same instruction, mode–route conditions and detached controller features through a non-modulating condition input. It uses an ordinary low-rank increment and trains with the corresponding L2R controller frozen. L2R retains its jointly trained checkpoint. At evaluation, both use the same task/round-specific controller weights and rules, applied to their own observations.

Table 3. Condition-aware LoRA with matched data and a shared controller.

TaskMethod90 trajectories180 trajectories270 trajectories
BowlsCondition-aware LoRA62% (31/50)66% (33/50)68% (34/50)
BowlsL2R76% (38/50)84% (42/50)88% (44/50)
TubesCondition-aware LoRA10% (5/50)18% (9/50)24% (12/50)
TubesL2R20% (10/50)28% (14/50)34% (17/50)

The three budgets contain 30/60/90 failure sources, with three recoveries per source. Evaluation retains the original nominal/recovery schedule of 30/5 steps. With more data, condition-aware LoRA improves on both tasks, yet L2R retains gaps of 14/18/20 points on Bowls and 10/10/10 on Tubes across R1/R2/R3. The advantage persists across all three rounds with matched teacher corrections, conditioning information, controller, and cumulative trajectory budgets.

W2(c): Weak baselines—“existing self-recovery or self-improvement frameworks.”

Response: We added a single-round HG-DAgger-style baseline [1]. A human operator corrects policy failures with a handheld teleoperation controller, and an ordinary LoRA student learns from these interventions. LoRA-BC and L2R instead share the tool-generated corrections.

Table 4. Intervention learning with 90 recovery trajectories per task.

MethodBowls SRTubes SR
Human-HG78% (39/50)20% (10/50)
LoRA-BC66% (33/50)12% (6/50)
L2R-R176% (38/50)20% (10/50)

Each method uses 30 failure boundaries with three recoveries each. All policies replan every five actions throughout the episode and deploy without an online expert. With the same tool-generated corrections and replanning interval, L2R exceeds ordinary LoRA-BC by 10 and 8 percentage points on Bowls and Tubes. L2R records one fewer success than Human-HG on Bowls and the same number on Tubes, approaching the human-intervention reference in these evaluations.

W2(d): Online tool-using agent—task success and end-to-end runtime.

Response: We evaluated Base assisted by GPT-5.6 Sol, the original RoboDojo collection teacher specified in Appendix A. The agent decides when to intervene and retains its perception tools, robot skills and verifier. L2R-R3 uses the original 30/5-step nominal/recovery schedule.

Table 5. Online recovery and policy-only deployment.

TaskSystemTask SRTime per episode (min)
BowlsL2R-R388% (44/50)3.2
BowlsBase + online teacher92% (46/50)6.8
TubesL2R-R334% (17/50)3.4
TubesBase + online teacher46% (23/50)8.4

End-to-end runtime is averaged over all 50 episodes per configuration. L2R reduces mean episode time by approximately 53% on Bowls and 60% on Tubes, with success rates 4 and 12 percentage points below online assistance. This provides a task-specific comparison of autonomous execution and online correction.

The tools and privileged simulator poses support training-data collection. Averaged across the two tasks, collection from a given failure to a successful recovery takes 30.4 minutes without tool assistance and 4.8 minutes with our tool-assisted pipeline. These are collection times, distinct from the from-start deployment times above.

AuthorsOpenReview ↗

Official Comment by Authors

W3: Policy-agnostic experiments—“different base policies and different reasoning agents.”

Response: We changed either the Base policy to π0.5 or the collection-time reasoning model to Opus 5.5 and evaluated one learning round. The original teacher is GPT-5.6 Sol. The teacher swap retains the XR-1 failure sources and tool interface; each Base policy learns from its own failures.

Table 6. Changing the Base policy or reasoning model.

TaskBaseTeacherBase SRR1 SRGain (pp)
BowlsXR-1gpt5.6-sol62% (31/50)76% (38/50)+14
Bowlsπ0.5gpt5.6-sol40% (20/50)56% (28/50)+16
BowlsXR-1Opus 5.562% (31/50)78% (39/50)+16
TubesXR-1gpt5.6-sol8% (4/50)20% (10/50)+12
Tubesπ0.5gpt5.6-sol6% (3/50)14% (7/50)+8
TubesXR-1Opus 5.58% (4/50)20% (10/50)+12

Each configuration uses 90 recoveries from 30 sources and the original execution schedule. Both Base policies improve on both tasks: π0.5 gains 16/8 points on Bowls/Tubes, while XR-1 gains 14/12. With Opus 5.5, XR-1 gains 16/12 points. Thus, the gains extend beyond the original policy–teacher pairing, with task- and policy-dependent improvements.

Q1: Full-policy evaluation under distribution shift.

Response: VLABench's shifted tracks supply recovery boundaries and conditions. We evaluated the full R1 policy from task start using RoboDojo's official Standard and Random task variants.

RoboDojo designates Standard for training and evaluation and Random for evaluation only. The official Random variants use different task-object instance or texture pools, with additional tabletop clutter in task-specific configurations. For example, Stack Bowls uses a separate bowl-instance pool with distractors, and Arrange Largest Number uses different digit textures.

Five tasks contribute 25 Standard and 25 Random episodes each: Stack Bowls, Pack Objects into Box, Store Laptop and Headphones, Stack Blocks, and Arrange Largest Number. Both conditions use the same R1 weights and thresholds, with autonomous entry, route selection and exit.

Table 7. Full-policy R1 evaluation with episode-level entry metrics.

ConditionEpisodesMacro task SREntry precision / recall
Standard12541.6% (52/125)91.1% / 96.8%
Random12520.8% (26/125)88.4% / 95.2%

Task SR weights the five tasks equally. Each episode counts once: a reference-positive episode needs recovery, and a predicted-positive episode contains at least one entry. Precision/recall are 92/101 and 92/95 for Standard, and 99/112 and 99/104 for Random.

With fixed weights and thresholds, most recovery-needing episodes contain an entry in both variants (recall 96.8%/95.2%). False triggers occur in 9/30 reference-negative Standard episodes and 13/21 Random episodes; observed false-trigger interventions return to nominal execution and remain counted. Task success is 41.6%/20.8%, so entry coverage and full-task completion are evaluated separately. Table 1 tests cross-task transfer of the controller–adapter pair.

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

This paper solves the problem of enabling VLAs to recover autonomously from execution failures by iteratively internalizing agent corrections. The method extracts effective recovery actions from successful interventions and learns when to enter recovery, how to select recovery behaviors, and when to exit recovery. The authors evaluate the conditional recovery capability and the success rate of execution from the task start on both simulated and real robot tasks, demonstrating the potential of iterative error-correction learning.

Critical strengths

This paper has a reasonable starting point, addressing the critical issue of execution failure and recovery in the real-world deployment of VLAs. The method simultaneously considers "how to recover" and "when to recover", and attempts to transform the error-correction capability of an external agent into a policy capability that does not require an online agent, which holds practical value. The experiments distinguish between recovery evaluation under given failure conditions and autonomous execution of complete tasks, and include simulations, real-machine experiments, and comparative studies on iterative data collection, providing solid preliminary support for the overall framework.

Critical weaknesses

The existing experiments have not yet sufficiently verified the trajectory recovery process during autonomous execution. The conditional recovery success rate and the final task success rate cannot separately illustrate whether the policy accurately identifies failures, whether it can recover to an effective execution state, and whether it can continuously complete the remaining tasks after recovery. In addition, a higher replanning frequency is adopted during the recovery phase, and part of the performance gains may come from denser feedback; therefore, the overall performance improvement cannot be fully attributed to the proposed recovery mechanism.

I suggest that the authors could consider reporting the recovery trigger accuracy, the recovery success rate for effective execution, and the remaining task completion rate after recovery separately under different failure types, occurrence positions, and severity levels. It would also be valuable to analyze whether failures occur during the entry, execution, or exit of the recovery phase. Furthermore, could a baseline strategy with the same high-frequency replanning be added to disentangle the gains brought by control frequency from those brought by the module itself?

Questions for authors

See above

4 comments
AuthorsOpenReview ↗

Official Comment by Authors

Thanks for raising these points about how recovery proceeds during autonomous execution. We have added the process analysis and matched-frequency comparisons to Appendix E of the revised manuscript.

W1 (Q1): Recovery entry, effective correction and subsequent completion.

Response: We analyzed autonomous R3 execution on the manuscript's four-task subset: Stack Bowls, Insert Tubes, Store Laptop and Headphones, and Fasten Screws. Each task contributes 50 episodes.

Table 1. Task success and event-level recovery metrics on four tasks at R3.

Evaluation setTask SREntry precision / recallEffective recoveryValid handoffCompletion after handoff
Four tasks, 200 episodes43.5% (87/200)89.9% (152/169) / 98.7% (152/154)46.1% (71/154)76.0% (73/96)76.7% (56/73)

Entry precision and recall use one-to-one matches between recovery entries and eligible failures defined by task-specific criteria. Effective recovery measures restoration to an executable state, with missed triggers retained in the denominator. Valid handoff requires an executable state at controller-initiated exit; completion after handoff requires Base to finish without another recovery.

Recovery entry covers 152 of 154 eligible failure events, while 71 are restored to an executable state. This distinguishes high entry coverage from the remaining challenge of executing effective corrections. After valid handoff, Base completes 56 of 73 cases without another recovery. This measures continuation from executable handoff states. The 23 exits outside executable states highlight the need for successful restoration before returning control to Base.

Observed false-trigger interventions return control to Base and remain in the entry and handoff statistics. Their handoffs count as valid only under the same effective-state criteria. Recovery that directly completes the task contributes to Task SR rather than handoff counts.

Failure definitions. Failure conditions are defined per task during teacher rollouts; routes specify corrective behaviors and context. Different failures can share a correction, and the appropriate correction can depend on the current state. We retain these task-specific definitions and analyze entry, restoration and handoff as common stages. Table 1 reports pooled results across failure types, occurrence positions and severities.

W2 (Q2): Contribution of higher replanning frequency.

Response: We compared the R1 LoRA-BC and L2R checkpoints at the same five-action replanning interval, using the same 90 recovery trajectories per task. We also evaluated the unchanged Base at prefixes 30 and 5.

Table 2. Task success at matched replanning frequency.

MethodExecuted prefixBowls SRTubes SR
Base3062% (31/50)8% (4/50)
Base564% (32/50)8% (4/50)
LoRA-BC-R1566% (33/50)12% (6/50)
L2R-R1576% (38/50)20% (10/50)

Prefix-5 policies replan every five actions throughout the episode. The Base comparison keeps the checkpoint, 30-action prediction horizon, initializations, native step limits and success criteria fixed.

Compared with prefix 30, Base at prefix 5 records one additional success on Bowls and the same number on Tubes. At prefix 5, L2R exceeds Base by 12 points on both tasks and LoRA-BC by 10 and 8 points: L2R's advantage remains at matched replanning frequency. VLABench comparisons also use a common five-action prefix.

A complementary R3 comparison in Appendix E (Table E1) uses the same 270 recovery trajectories per task. Both methods are evaluated with the same controller weights and the original 30/5-step nominal/recovery schedule. L2R reaches 88% on Bowls and 34% on Tubes, compared with 68% and 24% for condition-aware LoRA. The 20- and 10-point advantages persist at a larger correction-data budget under shared control.

Area_Chair_9ZBrOpenReview ↗

Reviewer: please reply

Please reply to the authors, and indicate whether your concerns were met.

Thanks for your response. I think most of my concerns have been addressed. Since I already gave a positive score, I'll currently keep my score.

Thank you for following up. We’re glad that our response addressed most of your concerns, and we appreciate your constructive suggestions again, which helped us strengthen the paper.

Reviewer_2pnVOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Learning to Recover (L2R), which gradually internalizes the failed correction data generated and verified by the reasoning agent into the autonomous recovery capability of the VLA model. This method utilizes conditioned LoRA to learn recovery actions and uses a lightweight local controller to determine when to enter, select, and exit the recovery mode; subsequent training rounds further collect and update new strategies for failure. Experiments show that L2R significantly outperforms standard LoRA on VLABench and continuously improves the success rate of tasks without the need for online agents.

Critical strengths
  1. L2R not only uses intelligent agents to address current failures, but also transforms one-time online corrections into reusable strategic capabilities, thereby reducing the continuous reliance on expensive inference models, manual intervention, and external tools during deployment.

  2. The paper models "how to perform recovery" and "when to invoke recovery" as conditionally low-rank adapters and local controllers, while keeping the basic VLA parameters frozen. This design can reduce interference to the original capabilities and support conditional switching between normal strategies and recovery strategies.

  3. The paper not only demonstrates the continuous benefits of multi-round training, but also shows through a controlled experiment that matching the data volume and update budget: collecting new failures of the updated strategies is more effective than repeatedly optimizing the existing data or continuing to collect failures of the basic strategies.

  4. L2R has been evaluated and validated for effectiveness on VLABench, ten RoboDojo tasks, and three real robot tasks.

Critical weaknesses
  1. The training requires reasoning agents, perception tools, robotic skills, and task validators to generate multiple valid recovery trajectories. The simulation data collection also uses privileged object poses. This results in the iteration requiring too many dependencies.

  2. The paper trains conditional adapters and local controllers separately for each task, but does not prove whether the same recovery module can be shared across tasks, instructions, or robots. Since the recovery routes and calling logic may depend on task definitions, the task generalization ability of the method needs to be verified.

  3. There is insufficient direct comparison with existing recovery methods. The main baselines focus on Full Fine-tuning, standard LoRA, and basic strategies, lacking unified experimental comparisons with other online recovery methods.

  4. The iterative benefits are still limited. The improvement from R2 to R3 on RoboDojo is small, and some tasks do not show further improvement, which makes the yield (improvement vs cost) brought by more iterations smaller.

Questions for authors

Refer to Weakness.

4 comments
AuthorsOpenReview ↗

Official Comment by Authors

Thank you for your constructive suggestions and for recognizing the value of reusable recovery. The additional results and analysis are in Appendix E of the revised manuscript.

W1 (Q1): Training-time dependencies.

Response: The agent, perception tools, robot skills and validators generate and verify recovery demonstrations during training. Privileged simulator poses are also confined to collection and verification. At deployment, the local controller and policy determine recovery entry, actions and handoff from observations (Section 3.4), without online agent calls.

The tools make offline collection more efficient. Averaged across Stack Bowls and Insert Tubes, collection from a given failed trajectory to a successfully completed recovery episode takes 30.4 minutes without tool assistance and 4.8 minutes with our pipeline: approximately one-sixth of the time.

W2 (Q2): Generalization beyond a fixed task.

Response: L2R learns task-specific recovery modules. On VLABench's four shifted tracks, conditional RecoverySR is 47.8% versus 29.0% for LoRA-BC with supplied boundaries and conditions. Table 1 tests cross-task transfer, freezing the R1 controller and adapter without target recovery training.

Table 1. Cross-task transfer of frozen R1 recovery modules.

SourceTargetBase SRTransferred SRGain (pp)Target-trained R1 SR
Stack BowlsStack Blocks10% (5/50)16% (8/50)+620% (10/50)
Stack BowlsBuild Tower0% (0/50)4% (2/50)+410% (5/50)
Insert TubesFill Pen Holder6% (3/50)12% (6/50)+616% (8/50)

Each target uses the same 50 initializations across configurations. Related-task transfer helps, adding 4–6 percentage points without target corrections; target-trained R1 remains 4–6 points higher.

L2R performs task-level recovery post-training on a Base policy adapted to the deployment robot. For another robot, the same workflow starts from a robot-compatible Base and learns from verified corrections on that platform. Direct reuse of a trained recovery module additionally depends on compatible observation and action representations and robot-specific execution dynamics.

W3 (Q3): Recovery baselines.

Response: We added a single-round HG-DAgger-style baseline, Human-HG: an operator corrects policy failures with a handheld teleoperation controller, and an ordinary LoRA student learns from those interventions. We also compare with LoRA-BC trained on the same tool-generated corrections as L2R.

Table 2. R1 comparison at a matched five-action replanning interval.

MethodBowls SRTubes SR
Human-HG78% (39/50)20% (10/50)
LoRA-BC66% (33/50)12% (6/50)
L2R-R176% (38/50)20% (10/50)

Each method uses 90 recovery trajectories per task from 30 failure boundaries, with three recoveries each. All students replan every five actions throughout the episode and deploy without an online expert. L2R exceeds LoRA-BC by 10 and 8 percentage points on Bowls and Tubes. It records one fewer success than Human-HG on Bowls and the same number on Tubes, approaching the human-intervention reference.

We further compared policy-only deployment with Base assisted by the original online correction agent.

Table 3. Policy-only execution and online recovery.

TaskSystemTask SRTime per episode (min)
BowlsL2R-R388% (44/50)3.2
BowlsBase + online teacher92% (46/50)6.8
TubesL2R-R334% (17/50)3.4
TubesBase + online teacher46% (23/50)8.4

The online teacher determines when to intervene and retains its tools and skills. Times average all 50 episodes per configuration. L2R reduces mean episode time by approximately 53% on Bowls and 60% on Tubes, with success rates 4 and 12 percentage points below online assistance.

W4 (Q4): Diminishing returns.

Response: The ten-task macro success rate increases from 22.4% at R1 to 25.8% at R2 and 27.4% at R3. The successive gains are 3.4 and 1.6 percentage points, with 90 additional recovery trajectories per task in each round.

Appendix C compares alternative uses of the update budget on four tasks. At R3, updated-policy failure collection reaches 43.5%, compared with 36.5% for Base-failure collection and 34.0% for continued training on the original data. The 7.0- and 9.5-point gaps at the same R3 update budget show that which failures are collected still matters as marginal gains diminish. The gap to online correction in Table 3 suggests that the current policy remains limited in learning and executing the available corrections. The π0.5 and Opus 5.5 comparisons in Appendix E (Table E4) show larger performance differences across Base policies than across teachers, consistent with policy capability being a limiting factor under the tested budget.

Area_Chair_9ZBrOpenReview ↗

Reviewer: please reply

Please reply to the authors, and indicate whether your concerns were met.

Reviewer_2pnVOpenReview ↗

Response to author

I appreciate the authors' detailed response, which addresses most of my concerns. Overall, I am happy to maintain my positive rating.

Thank you for reviewing our response and for your encouraging feedback. We’re glad that the additional results and clarifications addressed most of your concerns. Your suggestions helped us strengthen the paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Summary of Revisions and Additional Results

Thank you to the reviewers and chairs for the thoughtful feedback and time devoted to our paper. The discussion helped us strengthen the evidence for L2R: learning reusable recovery from verified corrections, with no online agent needed at deployment. The additional results and protocols are collected in Appendix E of the revised manuscript and marked in light blue.

1. The gains remain under matched baselines and replanning frequency.

On Stack Bowls and Insert Tubes, L2R-R1 reaches 76%/20% success versus 66%/12% for ordinary LoRA-BC using the same tool-generated corrections and five-action replanning. Increasing Base's replanning frequency alone adds only 2/0 percentage points. The separate HG-DAgger-style human-intervention reference reaches 78%/20% at the same trajectory budget and replanning interval.

We also added success-only fine-tuning and extended condition-aware LoRA to 90/180/270 recovery trajectories. At 270 trajectories, with matched conditioning and shared controller weights, L2R retains advantages of 20 and 10 percentage points. These comparisons support the recovery-learning design across data budgets and execution frequencies (Table E1).

2. Recovery is evaluated through entry, correction and subsequent completion.

On the four-task R3 subset (200 episodes), entry precision/recall is 89.9%/98.7%, covering 152 of 154 eligible failure events. Of these 154 events, 71 are restored to an executable state. Separately, Base completes 56 of 73 valid handoffs without another recovery. False triggers remain in the entry statistics; their exits count as valid handoffs when they satisfy the same executable-state criterion. The analysis distinguishes recognizing failures from correcting them and shows that valid handoffs support continued execution (Table E2).

3. Cross-task transfer and controller evaluation.

Frozen R1 controller-adapter pairs improve success by 4–6 percentage points across all three transfers: Stack Bowls → Stack Blocks, Stack Bowls → Build Tower, and Insert Tubes → Fill Pen Holder, without target-task recovery training. These gains suggest that recovery behaviors can transfer between tasks sharing similar manipulation skills, such as stacking or insertion (Table E3).

The full R1 policy is also evaluated on five tasks using RoboDojo's official Standard/Random variants, with unchanged weights and thresholds. Episode-level precision/recall is 91.1%/96.8% in Standard and 88.4%/95.2% in Random; task success is 41.6%/20.8%. The analysis separates coverage of recovery-needing episodes, false triggering and full-task completion (Table E8).

On four tasks at R3, the eight-query GRU reaches 43.5% macro success versus 38.5% for a current-query MLP. GRU histories of 4–16 queries lie within one point; the tested entry-threshold perturbations change macro success by at most 1.5 points. These results support the default history as a reasonable choice among the tested settings (Tables E5–E6).

4. The deployment benefit complements continued improvement.

On Bowls/Tubes, L2R-R3 reaches 88%/34% success versus 92%/46% for Base assisted online by GPT-5.6 Sol, while reducing mean episode time by approximately 53%/60%, without online agent calls (Table E7). Tools remain part of offline collection: from a given failure to successful recovery, they reduce the two-task mean collection time from 30.4 to 4.8 minutes.

As later-round gains diminish, the original four-task acquisition comparison still favors correcting updated-policy failures: R3 reaches 43.5%, versus 36.5% for Base-failure collection at matched accepted-trajectory and update budgets, and 34.0% for continued R1-data training at the same update budget (Appendix C). One-round gains also hold when replacing the Base policy with π0.5 or the teacher with Opus 5.5 (Table E4). Further progress can combine better learning and execution with broader recovery coverage, which more capable teachers could help provide.

We appreciate the reviewers' concrete suggestions. The revised evaluation makes the benefit, operating conditions and scope of reusable policy-only recovery clearer.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.