{"id":"eaa8a786-db0b-4397-a283-1df5912ff5ef","arxiv_id":"2608.26645","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FLARE endows Vision-Language-Action models with a retry/reset error recovery capability, achieving 84% average success rate across 9 robotic manipulation tasks.","lead":"This paper introduces FLARE, a framework that helps robots recover from mistakes during manipulation tasks. It uses a retry mechanism for small errors and a reset mechanism for bigger failures, guided by a vision-language model monitor, and was tested on 9 simulated and 2 real-world tasks with improved success rates.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Paper reports reset skill success rates as low as 20–24% (Table 4) but never measures how often those skills are invoked, so the 84.0% overall success rate may be fragile.","rationale":"The reader's weakest assumption focused on MLLM classification accuracy (88%). While that is a valid concern, I find a more acute vulnerability in the reset skill success rates themselves (as low as 20–24%) combined with the unknown frequency of their invocation. The reader noted reliance on simulation state for reset skill bootstrapping, which I agree is a limitation, but I emphasize the cascading risk: low reset skill performance plus imperfect MLLM plus unknown failure distribution makes the headline success rate poorly constrained. My proposed test would directly quantify this risk. The verdict remains CONDITIONAL because the concern is real but does not disprove the central claim—it merely reveals a missing piece of evidence.","tokens_in":18611,"tokens_out":3002,"duration_ms":28991,"concrete_test":"For each of the 9 tasks, record across the 50 evaluation trials: (i) how many OOD errors occur, (ii) which reset skill is invoked, (iii) the success/failure of each reset attempt. Then compute the overall success rate predicted by a probabilistic model that uses the reset skill success rates from Table 4 and the MLLM classification accuracy from Table 5, weighted by the observed invocation frequencies. Compare this predicted rate to the reported 84.0%. If the predicted rate is significantly lower (e.g., <75%), the reported success rate is not robust to the distribution of failure types encountered.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that FLARE achieves 84.0% average success across 9 tasks depends on the system's ability to handle OOD errors via learned reset skills. Table 4, however, reports that two of the four evaluated reset skills succeed only 20% (U-shaped block) and 24% (coffee pod). If these low-success skills are frequently triggered during evaluation, the overall success rate would be pulled down substantially. The paper provides no data on the frequency of OOD errors or which reset skills are used across the 50 evaluation trials per task. Without this, we cannot rule out that the 84.0% result is sustained only because the evaluation setup rarely encounters failures that require the difficult reset skills. Furthermore, reset skill bootstrapping depends on exact simulation state (MuJoCo logs) and on MLLM failure analysis (88% accuracy for reset object identification, Table 5). The compounding effect of low reset skill accuracy and imperfect MLLM classification is unanalyzed. This is a load-bearing gap: the claimed robustness to OOD errors is not adequately supported by the presented evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"This paper presents FLARE, a framework that augments Vision-Language-Action models (VLAs) with two recovery mechanisms: a 'Retry' capability, enabled by injecting perturbation-bridging sequences into training data to decouple robot pose from environment state, and a 'Reset' capability, where offline MLLM (Gemini-2.5-Pro) analysis of failure videos identifies out-of-distribution (OOD) states and bootstraps object-centric reset skills. At inference, an online MLLM monitor orchestrates between task execution and the learned reset skills. The approach is evaluated on 9 simulated RoboMimic tasks and 2 real-world tasks, reporting an average 84.0% success rate against baselines including Phoenix and π0.5.","tokens_in":18924,"tokens_out":2867,"duration_ms":27875,"significance":"If the claimed results hold, FLARE would represent a substantial step toward making VLAs robust to execution errors. The perturbation-bridging augmentation is a principled technique to break spurious pose-state correlations, and the MLLM-driven reset skill acquisition offers a scalable way to handle catastrophic failures. The empirical gains over strong baselines (e.g., +11.8% over π0.5 average, +26.2% over Phoenix) are noteworthy. However, several load-bearing aspects of the evaluation need strengthening before the conclusions can be fully accepted.","major_comments":[{"comment":"The paper reports reset skill success rates as low as 20% (U-shaped block) and 24% (coffee pod) in Table 4, yet never provides the frequency with which these skills are invoked during the 50 evaluation trials per task. Without this information, the 84.0% overall success rate in Table 1 may be fragile—if low-success reset skills are rarely triggered, the aggregate number is not representative of true robustness; if they are frequently triggered, the number is likely an overestimate. The authors must report, per task, the number of episodes where a reset skill was called and the per-episode success after each reset. A sensitivity analysis showing the range of plausible overall success rates given the measured reset skill accuracies and plausible invocation counts is also needed.","section":"§5.2, Table 4; §4, Table 1"},{"comment":"Reset skill training in simulation relies on privileged MuJoCo state logs to extract exact environment configurations at failure timestamps (Sec. 3.3: 'high-fidelity environment state ... from our logs'). While the real-world validation uses Any6D pose estimation to bypass this, the primary quantitative evidence (Table 1) comes from simulation where such privileged information is used. The authors should either (a) provide an ablation where reset skills are trained without privileged state (e.g., using only visual pose estimates in simulation) to demonstrate that the approach does not depend on ground-truth coordinates, or (b) explicitly quantify the gap in reset skill performance when using estimated vs. true poses, and discuss how this affects the transferability of the simulated results to real-world settings.","section":"§3.3; §4, Real-world Validation"},{"comment":"The offline MLLM failure analysis (Table 5) achieves 78-88% accuracy for reset object identification and 66-78% for timestamp identification. These errors propagate to the reset skill dataset and to the online monitor. The paper does not analyze the impact of MLLM misclassifications on overall system performance. At a minimum, the authors should simulate the end-to-end success rate under the measured classification error rates (e.g., by injecting label noise into the training pipeline or by reporting how many evaluation trials were negatively affected by an MLLM mistake). Without this analysis, the claim that MLLM 'bootstraps' reset skills in a reliable way is not supported.","section":"§5.2, Table 5; §5.2, Ablation"}],"minor_comments":[{"comment":"The description of the perturbation-bridging augmentation (lines 215-225) refers to 'linear interpolation for position and SLERP for rotation' but does not discuss how collisions or joint limits are handled when the bridging segment is kinematically interpolated. Clarify whether such invalid trajectories are filtered out or whether the model is expected to learn despite them.","section":"§3.2"},{"comment":"The y-axis label 'Data Generation Success Rate (%)' is ambiguous. Does this mean the fraction of generated trajectories that are kinematically valid, or the fraction that led to a successful simulation rollout? Please clarify in the caption.","section":"§5.1, Fig. 4"},{"comment":"The text states: 'We explicitly scope out unrecoverable situations, such as when an object is too large for the gripper to re-orient ... or when an object falls outside the robot’s reachable workspace.' This is an important limitation that should be mentioned in the main paper, preferably in the conclusion or in a limitations paragraph.","section":"Supplementary Material, §1"},{"comment":"The term 'Markovian' is used to describe the VLA policy (Sec. 3.1, line 140). However, the policy receives a visual observation that implicitly contains temporal context (e.g., object motion blur, changing scene). The architecture is not strictly Markovian; clarify what is meant.","section":"Throughout"},{"comment":"The notation 'd_i^B' is introduced but used inconsistently later in the paragraph (e.g., 'd_B' appears without the subscript). Standardize the notation.","section":"§3.2"}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Here's my read on FLARE. The paper identifies a real weakness in current VLAs—they break on small deviations because training data is trajectory-monotonic—and proposes a sensible two-part fix: perturbation-bridging augmentation for in-distribution retries, and an MLLM-bootstrapped reset skill library for out-of-distribution failures. The ID/OOD error taxonomy is useful framing. The simulation results on 9 tasks are extensive and the average 84% is impressive relative to baselines. They also attempt real-world validation on two tasks, which is more than many simulation-only papers do.\n\nNow the soft spots. The stress-test concern is valid: Table 4 reports that two of the four reset skills succeed only 20–24% of the time, but the paper never tells you how often those skills are triggered during the 50 evaluation trials. If the evaluation setup rarely encounters failures that demand the hard resets, the headline 84% could be sustained by luck or by a forgiving test distribution. Without that breakdown, the claimed OOD robustness isn't fully supported. Also, the MLLM monitor hits 88% accuracy on reset-object identification (Table 5); the paper doesn't analyze what happens when it misclassifies—does that cascade into a full task failure? The real-world validation is thin (only two tasks, with limited object variation). And while they replace MuJoCo state with Any6D pose estimation, they don't quantify how well that works under the conditions of their experiments.\n\nCredit where due: the perturbation-bridging idea is clever and the ablations on its hyperparameters are thorough. The authors acknowledge some limitations (hardware constraints, need for larger reset libraries).\n\nWho should read this? Anyone working on making VLA policies robust enough for deployment. It's a serious piece of work with a clear narrative and honest limitations.\n\nFor peer review: this paper deserves a shot at a venue like CoRL or ICRA. I'd recommend conditional acceptance requiring (a) a breakdown of OOD error frequency and reset-skill usage per trial, and (b) either code release or a third real-world task. The central paradigm is sound; the missing data can be added without re-running the entire study.","headline":"FLARE tackles the right problem (VLA brittleness) with a clean retry/reset paradigm and solid simulation results, but the evidence for OOD recovery is thinner than advertised: reset-skill success rates as low as 20–24% are reported without showing how often those skills are actually invoked.","tokens_in":19447,"tokens_out":1654,"duration_ms":17188},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"FLARE claims that VLA brittleness comes from pose-state correlation learned from imitation data, and shows that retry plus reset training lifts average manipulation success to 84 percent over nine RoboMimic tasks.","keywords":["vision-language-action models","robotic manipulation","error recovery","retry-reset paradigm","perturbation-bridging augmentation","multimodal LLM failure analysis","imitation learning","robot self-correction"],"falsifier":"Compare the full FLARE pipeline on a fixed task under two labeling regimes: automatic MLLM failure labels and human ground-truth labels, with all training data and augmentation held constant; if the two land within a few points, the MLLM analyst is not the bottleneck, while a large gap would show the reported gain depends on the analyst rather than on the retry/reset data mechanism. A second direct check is to remove bridging segments and train only on the perturbed poses; if retry success collapses, bridging is the component that teaches recovery, whereas if it survives, pose diversity alone explains the result.","tokens_in":20366,"feed_emoji":"🤖","tokens_out":12556,"duration_ms":107470,"temperature":0.7,"pith_summary":"The paper identifies the brittleness of vision-language-action models (VLAs) as a data problem: human demonstrations follow a narrow pose path and almost never show recovery, so the policy learns a spurious link between its own arm pose and task progress. FLARE addresses that root cause with a two-part retry-and-reset mechanism that separates simple pose deviations, which a retry can absorb, from catastrophic state changes, which need an object reset. Retry is built into the policy by injecting random pose perturbations and bridging segments between subtask demonstrations, so the model learns to execute the correct next action from many arm poses. Reset is bootstrapped by a multimodal large language model (MLLM) that mines failure videos for labels, which seed a small library of object-centric recovery demonstrations, each stored as a separate adapter on the same backbone. The paper reports average success rising from 72.2 percent for the base VLA to 84.0 percent across nine manipulation tasks, with a two-task real-world check moving the same way; if the mechanism is what those numbers suggest, recovery capability can be engineered into training data rather than bolted on at deployment.","feed_headline":"Retry/reset training lifts robot task success to 84 percent","feed_subtitle":"A training recipe that decouples robot pose from scene state, plus learned reset skills, lets VLA agents fix their own mistakes.","key_machinery":"The load-bearing machinery is a retry/reset policy library built on an in-distribution (ID) versus out-of-distribution (OOD) error taxonomy. ID errors are handled by perturbation-and-bridging augmentation: before each subtask segment the arm is moved to a random pose, a kinematically interpolated bridging segment returns it to a valid starting pose, and the model trains on bridging-to-task subsequences so the correct next action no longer depends on the arm's pose. OOD errors are handled by a reset pipeline: an MLLM failure analyst converts execution videos into structured labels (retry or reset, the object to reset, the interacting error group, and the failure timestamp), and those labels bootstrap a compact set of object-centric reset demonstrations, expanded with the same augmentation and stored as separate LoRA adapters. An online MLLM monitor closes the loop by swapping between the task adapter and the matching reset adapter, and back again when the environment is valid.","core_discovery":"FLARE's central claim is that a single imitation-trained VLA can be given autonomous recovery by changing what it is trained on, rather than by attaching a separate corrective controller. The root cause it names is that demonstrations are trajectory-monotonic and success-biased: the robot follows a narrow manifold of poses tightly correlated with task progress, and the policy overfits to that correlation. The paper's taxonomy splits the resulting failures into in-distribution errors, where the environment is still valid but the arm is in an unseen pose, and out-of-distribution errors, where the environment itself is broken, such as a toppled cup. Perturbation-and-bridging augmentation inserts random arm motions and smooth return segments between subtasks and trains on the return-plus-task sequences, teaching the policy to act from any arm pose. An offline MLLM then labels failure videos to identify which object needs resetting, those labels bootstrap roughly twenty human demonstrations per object-centric reset skill, the same augmentation expands them, and each skill is stored as a LoRA adapter on the shared pi0.5 backbone. At inference an online MLLM monitor leaves retry-capable behavior alone and swaps in the reset adapter when a state-breaking failure appears, then resumes the task. The reported result is 84.0 percent average success across nine RoboMimic tasks, 11.8 points over the pi0.5 backbone and 26.2 points over the Phoenix self-reflection baseline, and the system also beats Phoenix-Human, the same baseline given correct human correction.","pith_inferences":["If trajectory monotonicity is the root cause, the perturbation-and-bridging recipe should transfer to other imitation-learned policies, including diffusion policies, with no architectural change.","The reset loop generates its own failures, so the same MLLM mining could be run recursively to grow a second-generation reset library, turning recovery into a hierarchy of increasingly specific skills.","The two real-world tasks and 40 trials per task are too few to establish that the simulation gains survive in cluttered or deformable-object settings; that is a testable claim rather than a demonstrated one.","Reset-skill success rates are far lower for pose-critical objects (24 percent for the coffee pod and 20 percent for the U-shaped block), suggesting the next bottleneck is dexterous in-hand pose correction, not failure detection or task arbitration."],"forward_implications":["A VLA trained with perturbation-and-bridging augmentation can absorb minor pose deviations during execution without any external correction signal, eliminating many resets at the source.","Only ten demonstrations per task and twenty per reset skill are enough, when augmented, to build both the main policy and the recovery library.","New recovery behaviors can be added after deployment by training a new LoRA reset adapter, without retraining or disturbing the existing task policy.","The largest gains appear on harder randomization settings, consistent with the claim that decoupling pose from environment state is what drives robustness to variation.","Recovery from state-breaking failures depends on a finite library of object-centric reset skills, so failures outside the library remain unrecoverable until a new skill is added."],"supporting_citations":[{"why":"is the pre-trained VLA backbone that all task and reset adapters fine-tune, and the main performance baseline that FLARE improves on.","marker":"[15]"},{"why":"supplies the demonstration expansion and planner-free kinematic interpolation used to build trajectories and perturbation-and-bridging segments.","marker":"[29]"},{"why":"provides the self-reflection baseline, task suite, and evaluation protocol that the comparisons are measured against.","marker":"[39]"},{"why":"provides the low-rank adaptation mechanism that lets each task and reset skill be trained as a separate adapter on the shared backbone.","marker":"[14]"},{"why":"is the multimodal LLM used as the offline failure analyst and online monitor for reset-skill mining and arbitration.","marker":"[34]"},{"why":"provides the RoboMimic simulation tasks used in the nine-task evaluation.","marker":"[30]"},{"why":"provides object pose estimation that lets the real-world experiments perform augmentation from visual sensors instead of privileged simulator state.","marker":"[19]"}],"fun_headline_variants":["VLA robots self-correct with retry and reset, 84% success","Decoupled pose and scene training lets robots fix own errors","Failure-aware training boosts robot task success to 84%"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The load-bearing premise is that a vision-language model watching execution video can reliably tell a retryable deviation from a state-breaking failure and can name the object that must be reset; mislabels would make the system skip needed resets or run unnecessary ones, and the paper does not trace how often that happens in the final success numbers.","fun_headline_variants_meta":{"raw":{"variants":["VLA robots self-correct with retry and reset, 84% success","Decoupled pose and scene training lets robots fix own errors","Failure-aware training boosts robot task success to 84%"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000523,"raw_usage":{"cost":0.000522928,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000522928,"upstream_inference_prompt_cost":0.000127848,"upstream_inference_completions_cost":0.00039508},"total_tokens":2529,"prompt_tokens":1118,"completion_tokens":1411,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":1352}},"tokens_in":1118,"tokens_out":1411,"cost_source":"openrouter_usage","duration_ms":13486,"temperature":1.0,"reasoning_tokens":1352,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-09-09T09:50:56.590582+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-alltime-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Compare the full FLARE pipeline on a fixed task under two labeling regimes: automatic MLLM failure labels and human ground-truth labels, with all training data and augmentation held constant; if the two land within a few points, the MLLM analyst is not the bottleneck, while a large gap would show the reported gain depends on the analyst rather than on the retry/reset data mechanism. A second direct check is to remove bridging segments and train only on the perturbed poses; if retry success collapses, bridging is the component that teaches recovery, whereas if it survives, pose diversity alone explains the result.","supporting_citations":[{"cited_title":"$\\pi_{0.5}$: a Vision-Language-Action Model with Open-World Generalization","cited_arxiv_id":"2504.16054","evidence_quote":"is the pre-trained VLA backbone that all task and reset adapters fine-tune, and the main performance baseline that FLARE improves on."},{"cited_title":"Mimicgen: A data generation system for scalable robot learning using human demonstrations","cited_arxiv_id":null,"evidence_quote":"supplies the demonstration expansion and planner-free kinematic interpolation used to build trajectories and perturbation-and-bridging segments."},{"cited_title":"Phoenix: A motion-based self-reflection framework for fine- grained robotic action correction","cited_arxiv_id":null,"evidence_quote":"provides the self-reflection baseline, task suite, and evaluation protocol that the comparisons are measured against."},{"cited_title":"Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022","cited_arxiv_id":null,"evidence_quote":"provides the low-rank adaptation mechanism that lets each task and reset skill be trained as a separate adapter on the shared backbone."},{"cited_title":"Gemini: A Family of Highly Capable Multimodal Models","cited_arxiv_id":"2312.11805","evidence_quote":"is the multimodal LLM used as the offline failure analyst and online monitor for reset-skill mining and arbitration."},{"cited_title":"What matters in learning from offline human demonstrations for robot manipulation","cited_arxiv_id":null,"evidence_quote":"provides the RoboMimic simulation tasks used in the nine-task evaluation."},{"cited_title":"Any6d: Model-free 6d pose estimation of novel objects","cited_arxiv_id":null,"evidence_quote":"provides object pose estimation that lets the real-world experiments perform augmentation from visual sensors instead of privileged simulator state."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}