{"id":"15fccc94-0e7a-4da5-8f49-0f2913d01256","arxiv_id":"2607.13818","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":15,"one_line_summary":"A learned high-level agentic policy that selects retry/repair/reset modes on top of frozen VLA policies improves LIBERO success rates by up to 13.7% (standard) and 39.2% (disturbance) over baselines.","lead":"The paper wraps frozen robot manipulation policies in a learned 'agentic' manager that watches execution quality and can trigger retry, repair, or reset modes. On the LIBERO benchmark it reports higher success rates, especially under injected disturbances.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No ablation against fixed recovery policies: the reported gains may reflect recovery primitives rather than learned agentic decisions.","rationale":"The reader's weakest assumption—representativeness of the 50-successful-trajectory reference library—is reasonable but less central. In the reported baselines, even the worst nominal success rate (Diffusion Policy Long, 50.5%) makes collecting 50 successful trajectories feasible, and the observation in Sec. IV-C.1 includes raw proprioception and low-level actions, so the agentic policy could partially compensate for a noisy global-quality metric. The missing ablation against fixed recovery policies is more damaging because it determines whether the central 'learning' contribution is real or whether the gains are an artifact of adding recovery primitives. The paper is internally consistent, but the claim that the policy 'learns to restore effective execution' requires evidence that learning adds value over hand-coded or trivial policies. This does not overturn the CONDITIONAL verdict, but acceptance should require such an ablation.","tokens_in":11173,"tokens_out":7693,"duration_ms":76441,"concrete_test":"On LIBERO-Long with OpenVLA and Diffusion Policy under the same disturbance protocol, evaluate three fixed high-level policies over at least 100 episodes per condition with 3 seeds: (a) hand-coded thresholds: RETRY when q_agg<0.5 for 3 consecutive decisions, REPAIR when contact force persists for 5 steps, RESET when q_agg<0.3 for 10 steps; (b) simplest policy: RETRY once after the first q_agg drop by 0.2, then EXECUTE; (c) uniform-random mode selection as a control. Compare success rates to the trained agentic policy. If (a) or (b) comes within 5 percentage points absolute of the trained policy, the learned decision-making is not established; if the trained policy beats (a)/(b) by a wide margin and (c) performs poorly, the concern is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a PPO-trained agentic policy learns to regulate execution and improves robustness. But the only experimental comparison is Table II's 'w/o Agentic' vs 'w/ Agentic'. The 'w/ Agentic' condition bundles two interventions: (i) the execution-quality metrics plus fixed recovery mechanisms (RETRY/REPAIR/RESET, Sec. IV-B) and (ii) the learned policy that selects when to invoke them (Sec. IV-C). There is no baseline that uses the same metrics and recovery mechanisms with hand-coded or fixed triggering rules. For example, a threshold policy that issues RETRY whenever q_agg falls below a threshold, REPAIR when sustained contact is detected, and RESET after prolonged low quality would isolate the value of learning. If such a fixed policy recovers most of the +13.7%/+39.2% gains, then the learned selection is epiphenomenal and the contribution reduces to adding recovery primitives. Sec. V.A says baselines are 'without execution-level decision-making', but that only excludes no-decision executers; it does not compare against non-learning decision policies. The internal analyses (Figs. 6-7, Table III) demonstrate that the learned policy behaves plausibly, but they do not quantify the performance gain attributable to learning over a fixed schedule. This missing control directly affects the central 'learning' claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an execution-level wrapper for frozen low-level manipulation policies. It introduces a local execution quality metric (Eqs. 1-4) and a global execution quality metric based on a reference library of successful trajectories, aggregates them into q_agg (Eq. 5), and trains a high-level PPO policy over four discrete modes: EXECUTE, RETRY, REPAIR, and RESET. The agentic policy observes a history of proprioception, low-level actions, and quality scores, and triggers recovery mechanisms without modifying the frozen policy. Experiments on LIBERO with OpenVLA, π0, π0.5, and Diffusion Policy report success-rate improvements up to +13.7% average under standard conditions and +39.2% under injected disturbances.","tokens_in":11639,"tokens_out":5675,"duration_ms":52901,"significance":"If the learned agentic selection is indeed the cause of the reported gains, the approach would be a lightweight, policy-agnostic mechanism for improving robustness of VLAs and diffusion policies without retraining, with interpretable decisions and modest overhead. The paper addresses a real gap in current VLA deployment and its design is sensible. However, the evidence as presented bundles the fixed recovery mechanisms with the learned policy, lacks statistical uncertainty measures on the main table, and leaves several load-bearing hyperparameters unspecified. The significance is therefore conditional on addressing these empirical gaps.","major_comments":[{"comment":"The 'w/o Agentic' vs 'w/ Agentic' comparison bundles two interventions: (i) the execution-quality metrics plus fixed recovery primitives (RETRY/REPAIR/RESET, §IV-B) and (ii) the learned PPO policy that selects when to invoke them (§IV-C). No baseline uses the same metrics and recovery mechanisms with fixed or hand-coded triggering. A threshold policy — e.g., RETRY when q_agg drops below a threshold, REPAIR when contact is sustained, RESET after prolonged low quality — is needed to isolate the value of learning. Without such a control, the reported +13.7/+39.2 gains may be due to the recovery primitives alone, which would weaken the central 'learning' claim.","section":"§V-A, Table II"},{"comment":"The main results are single point estimates with no error bars, number of seeds, or significance tests. Figure 5 provides 95% CIs only for one training run (LIBERO-Spatial/OpenVLA). The text's claim of 'consistent improvements' is contradicted by the π0.5 rows: 97.4→96.6 (Spatial) and 98.0→97.4 (Goal) under standard settings. Without variability information, the reader cannot tell whether these are noise or genuine degradations. All rows of Table II should be repeated over seeds and reported with CIs, or the claim must be revised.","section":"Table II, §V-B"},{"comment":"The paper states 'episodes in which RESET is invoked are counted as failures to ensure fair comparison.' This means an episode that invokes RESET and subsequently completes the task is scored as a failure, so the reported success rate is not simply task completion. This unusual criterion can distort the comparison and makes the metric less interpretable. Please report both the standard task-completion rate and the RESET-invoked rate, or justify why the current definition is appropriate.","section":"§V-A, RESET counting"},{"comment":"The global quality signal depends on a reference library of N=50 successful trajectories per task and on hand-set parameters B=10, k=5, α, λ, W, c, w1, w2, b, β, τ, M, N. No sensitivity analysis is provided. If the frozen policy cannot reliably produce 50 successful trajectories for a task, or if these thresholds do not transfer, the global quality signal is undefined. The paper should report the actual numerical values used and include ablations over at least N, B, k, α, and λ. Without this, the transferability claim in §VI is not supported.","section":"§IV-A.2, Eqs. (4)-(5)"},{"comment":"The text says 'The agentic policy is trained separately for each task' but also 'Training spans multiple tasks within each LIBERO subset.' These statements conflict. If the agentic policy is trained and evaluated on the same task, the comparison is in-distribution; if one policy is trained across tasks and evaluated on held-out tasks, the experimental protocol should state this explicitly and avoid any risk of leakage. Please clarify the exact train/test split for each row in Table II.","section":"§V-A, training/evaluation"}],"minor_comments":[{"comment":"Typo in affiliation: 'Inteligence' should be 'Intelligence'.","section":"Author affiliation"},{"comment":"The table formatting has minor issues: 'A vg.∆' and the use of shaded rows should be cleaned up for clarity.","section":"Table II"},{"comment":"The 'execution degradation types' are described as manually inspected and categorized. Please define the taxonomy and the criteria used, otherwise the conditional distribution in Fig. 6 is not reproducible.","section":"Fig. 6"},{"comment":"The 'inverse coefficient of variation' definition is unusual: S = 1/(1+σ²/μ²). Please clarify the relationship to the standard coefficient of variation and why this particular bounded form is used.","section":"§IV-A, Eq. (3)"},{"comment":"It is stated that information generated during recovery phases is not recorded in the execution history. Please specify whether the history length L=20 refers to low-level steps before the recovery or after, and how the decision interval K interacts with recovery phases.","section":"§IV-C"},{"comment":"Reference [15] uses the notation 'π∗0.6', which is inconsistent with the π0/π0.5 notation used elsewhere. Please standardize.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses a relevant problem and the proposed framework is plausible, but the empirical evidence is currently insufficient: the missing fixed-recovery baseline is the main obstacle, and the lack of error bars in the central table is a serious reporting gap. I believe both are fixable within the scope of a major revision, so I do not recommend rejection. I would advise the editor to request the additional baseline and a statistical reporting revision before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a plausible but incremental wrapper that puts a PPO-trained execution manager on top of frozen VLA policies, and the headline 'learning' claim is not yet fully supported because there is no control that separates learned selection from the recovery primitives themselves.\n\nCredit where due: the package is coherent and the components fit together. The local quality metric (effectiveness and smoothness) and the global metric (comparison against a stage-aware library of 50 successful trajectories) are reasonable choices, and the POMDP formulation is clean. The internal analyses — decision distribution conditioned on degradation type, the separation of modes in quality space, and the event-conditioned quality changes in Table III — suggest the learned policy is doing something meaningful rather than random triggering. The gains on OpenVLA and Diffusion Policy under disturbance are large and plausible. They also state plainly in the conclusion that severe OOD degradation remains a limitation, which I take as honest.\n\nThe main soft spot is the missing baseline. The 'w/ Agentic' condition bundles the quality metrics, the fixed recovery mechanisms, and the learned policy. Without a fixed rule-based recovery policy (e.g., RETRY when q_agg is low, REPAIR on sustained contact, RESET after prolonged low quality), the +13.7%/+39.2% gains might come mostly from the recovery routines, with PPO adding little. Table III and Figs 6-7 soften this concern but don't quantify the learning contribution. This is the gap I'd want closed before the central claim is accepted.\n\nSmaller issues: Table II has no error bars or significance tests; the phrase 'without degradation' is contradicted by the pi0.5 cells (Spatial 97.4→96.6, Goal 98.0→97.4). There are many hand-set hyperparameters (W, c, b, k, α, λ, M, N, τ, K, L, costs, bin count, reference-library size) with no sensitivity analysis. No code or data, which matters because the global metric requires collecting 50 successful trajectories from the frozen policy — that's an empirical feasibility assumption that isn't discussed for weaker policies. RESET counted as failure is fair, but the paper reports no reset frequency, so we can't see how much of the gain is restart-driven.\n\nFor peer review: yes, it deserves review, but I'd want the artifact and the fixed-recovery ablation before believing the learning part of the claim. The engineering package is useful and the paper is readable. I'd bring it to reading group for a discussion of evaluation norms.","headline":"Plausible but incremental execution wrapper for frozen VLA policies; the learning claim needs a fixed-recovery baseline before it holds.","tokens_in":12030,"tokens_out":3201,"would_cite":false,"duration_ms":33399,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that a frozen manipulation policy can be made resilient to execution-level failures by adding a lightweight learned manager that chooses among four recovery modes, raising success rates by up to 13.7% nominally and 39.2% un","keywords":["execution-level failure","agentic reinforcement learning","vision-language-action models","robotic manipulation","recovery mechanisms","execution quality metrics","LIBERO benchmark","hierarchical decision-making"],"falsifier":"Train the agentic wrapper on a task where the frozen policy succeeds in fewer than, say, 10 of 50 attempts, so the reference library is thin or biased; if the method requires the full 50 to trigger recovery, its gains should shrink to the baseline. Alternatively, inject a disturbance type that the recovery mechanisms do not cover — for instance, an object dropped out of the workspace — and check whether the success rate with the agentic policy remains close to the baseline, indicating the recovery ladder is not actually selecting appropriate responses.","tokens_in":11122,"feed_emoji":"🤖","tokens_out":3977,"duration_ms":39529,"temperature":0.7,"pith_summary":"The paper argues that execution-level failures in frozen manipulation policies can be detected and corrected by a lightweight agentic policy trained with reinforcement learning, which selects among four recovery modes. Rather than generating new low-level actions, this manager observes a short history of proprioception, actions, and two runtime quality scores and decides whether to continue, retry, repair, or reset. On the LIBERO benchmark it raises success rates by up to 13.7% under standard conditions and up to 39.2% when random kinematic disturbances are injected, without retraining the frozen policy. The key insight is that execution management and action generation can be cleanly separated, turning a purely conditional action generator into a decision-driven execution framework.","feed_headline":"Agentic RL lifts success up to 39% under disturbances","feed_subtitle":"A lightweight policy choosing EXECUTE, RETRY, REPAIR, or RESET boosts robustness without retraining the base model.","key_machinery":"The load-bearing object is a POMDP formulation of execution management, with a discrete action space A={EXECUTE, RETRY, REPAIR, RESET}. The agentic policy receives an execution history (the last L steps of proprioception, low-level actions, and q_local/q_global) and is trained with PPO. The quality metrics bridge raw execution and decision: q_local aggregates motion effectiveness (distance moved divided by commanded action magnitude) and smoothness (inverse coefficient of variation of speed); q_global compares a feature vector of the recent window against k-nearest neighbors in progress-binned reference trajectories. Recovery is realized by an operational space controller that returns the ro","core_discovery":"The paper's central claim is that execution-level degradation in robotic manipulation can be detected and corrected by a high-level agentic policy trained with PPO, which regulates a frozen low-level policy. Rather than generating new actions, the agentic policy selects among four execution modes, using two metrics: local execution quality (motion effectiveness and smoothness in a sliding window) and global execution quality (distance of the current trajectory prefix to a stage-aware library of successful reference trajectories). When quality drops, RETRY rolls back to the best recent state, REPAIR rolls back to a contact-free state further in the past, and RESET restarts the episode. The au","pith_inferences":["The stage-aware reference library suggests a natural generalization: the same rollout library could be built online during deployment, letting the system adapt to new tasks without a separate collection phase.","The binary contact indicator (threshold 5 N) hints that the rollback selection could be transferred to real hardware using any contact-force proxy, but the hand-set parameters (B=10, k=5, alpha, lambda, M, N) would likely need per-domain tuning.","A testable extension is to replace the hand-designed q_local with a learned critic of the same variables, potentially removing the manual sensitivity constants while retaining the recovery structure.","The framing as execution management rather than replanning implies a separation that could also apply to non-manipulation domains like mobile navigation, where a frozen planner could be regulated by a similar agentic loop."],"forward_implications":["The framework converts failure-prone policies (OpenVLA, diffusion) into substantially more reliable executors, with the largest gains on LIBERO-Long, the most compositionally difficult suite.","Because the low-level policy stays frozen, the approach can wrap any existing manipulation policy without retraining it or collecting new demonstrations.","Recovery decisions are interpretable: mild degradations trigger RETRY, collisions trigger REPAIR, and irrecoverable states trigger RESET, and each decision measurably improves subsequent execution quality.","Overhead is modest — roughly one to two extra recoveries and 5–15% longer episodes, scaling with the fragility of the base policy."],"fun_headline_variants":["Agentic RL picks EXECUTE, RETRY, REPAIR, RESET to lift success 39%","39% more success in disturbed manipulation via agentic RL","Agentic policy regulates execution to improve robot success 39%","Recovery modes from agentic RL boost robot success under disturbance","Frozen policy + agentic RL yields 39% robustness gain"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The global quality score relies on a pre-collected library of 50 successful trajectories per task; if such trajectories cannot be obtained for a task, or the fixed binning and distance thresholds are not representative, the agentic policy cannot detect drift and the recovery gains collapse.","fun_headline_variants_meta":{"raw":{"variants":["Agentic RL picks EXECUTE, RETRY, REPAIR, RESET to lift success 39%","39% more success in disturbed manipulation via agentic RL","Agentic policy regulates execution to improve robot success 39%","Recovery modes from agentic RL boost robot success under disturbance","Frozen policy + agentic RL yields 39% robustness gain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000275,"raw_usage":{"total_tokens":1463,"prompt_tokens":710,"completion_tokens":753,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":454,"completion_tokens_details":{"reasoning_tokens":656}},"tokens_in":454,"tokens_out":753,"duration_ms":7297,"temperature":1.0,"reasoning_tokens":656,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T03:35:45.928842+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the agentic wrapper on a task where the frozen policy succeeds in fewer than, say, 10 of 50 attempts, so the reference library is thin or biased; if the method requires the full 50 to trigger recovery, its gains should shrink to the baseline. Alternatively, inject a disturbance type that the recovery mechanisms do not cover — for instance, an object dropped out of the workspace — and check whether the success rate with the agentic policy remains close to the baseline, indicating the recovery ladder is not actually selecting appropriate responses.","supporting_citations":[],"review_version":1}