{"id":"a69696fb-5e2b-42b5-a019-677f47cbe526","arxiv_id":"2607.08877","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Human corrective actions can be inverted into noise-space targets that train a lightweight latent policy to steer frozen flow/diffusion robot models from a handful of interventions while preserving pretrained skills.","lead":"FlowDAgger adapts frozen generative robot policies by mapping human corrective actions into the policy’s noise space and training a small steering network on those inverted targets. It offers a practical way to fix real-world failure modes of robot foundation models without expensive fine-tuning or unsafe online RL.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the paper's own manifold-support limit already flagged by the reader.","rationale":"The reader's weakest_assumption correctly isolates the manifold-support condition that §6 already states. Within that regime the evidence is multi-family (π0.5, Cosmos-Policy, Gr00t, diffusion), multi-platform, and baseline-controlled. Real-robot lack of seeds and absent code justify CONDITIONAL rather than unconditional ACCEPT; nothing stronger is required. The concrete_test is a cheap diagnostic that would confirm the support assumption is satisfied for the published results, not a new threat to the argument.","tokens_in":14622,"tokens_out":445,"duration_ms":4769,"concrete_test":"On one real task (e.g., Toolbox Packing), log per-correction reconstruction error ||π_gp(s, w*) − a*|| after inversion; if median error stays ≪ action scale and success still rises as in Table 4, the support assumption holds for the reported regime. If many corrections invert poorly yet headline gains remain, the mechanism would need re-examination.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim holds under the conditions the paper states. Action inversion (per-step fixed-point Eq. 3 for flow heads; joint EDM inversion for WAMs) plus supervised noise-policy regression on inverted corrections and successful autonomous noise is a coherent, well-supported adaptation interface. Tables 1–4 and Figs. 3–6 show consistent gains over SFT, residual/LoRA DAgger, and DSRL under matched budgets, with prior preservation on held-out tasks. The only load-bearing bound is the one already named in §6 and by the reader: desired a* must lie sufficiently inside the support of the frozen map so that π_gp(s, w*) ≈ a*. Reported experiments operate inside that regime (low Action MSE in App. B.1; real-robot improvements from 5–20 interventions). No internal inconsistency, hidden assumption, or unacknowledged failure mode undermines the claim within the paper's stated scope.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"FlowDAgger adapts frozen generative robot policies (flow-matching and diffusion action heads, and world-action models) from human interventions by mapping each corrective action a* at observation s to a noise vector w* such that the frozen base map satisfies π_gp(s, w*) ≈ a* (action inversion via per-step fixed-point reverse Euler for few-step ODEs, and joint EDM inversion with action-frame minimal-delta targets for WAMs). A lightweight noise policy is then trained by supervised regression on inverted corrections plus successful autonomous noises (dual buffer), and replaces the base noise draw at deployment without updating base weights. Experiments on MetaWorld (π0.5 and Cosmos-Policy), Gr00t N1.7, a diffusion policy, and eight real single-arm/bimanual tasks report higher success under matched budgets than SFT, LoRA-DAgger, Residual-DAgger, and DSRL, with better held-out prior preservation than weight-space fine-tuning.","tokens_in":14920,"tokens_out":1253,"duration_ms":16311,"significance":"If the results hold, the paper offers a practical, compute-light interface for adapting robot foundation models in the real world: few human interventions, consumer-GPU training, and no weight updates that erode pretrained skills. The combination of inversion-based noise targets with DAgger-style collection is a clear methodological contribution relative to residual action-space DAgger and reward-driven latent RL (DSRL). Strengths include multi-family transfer (action-head VLAs and WAMs), explicit inversion accuracy ablations (App. B.1–B.2), prior-preservation measurements (Table 3), and real-hardware gains from 5–20 intervention episodes (Table 4). The manifold-support limit is stated honestly in §6.","major_comments":[{"comment":"Table 4 (real hardware) reports only Base, SFT, and FlowDAgger. The central claim that noise-space adaptation outperforms action-space residual and weight-space DAgger is supported in simulation (Table 1, Fig. 3) but not isolated on hardware, where Residual-DAgger and LoRA-DAgger are absent. Because residual methods are the closest architectural competitors and can be run with the same intervention stream, their omission leaves the real-world advantage of the latent interface under-supported. Adding at least Residual-DAgger (or a residual action-space control) under the same intervention budget would make the hardware claim load-bearing rather than suggestive.","section":null},{"comment":"§4.2 and App. A claim that joint world-action inversion with a minimal-delta action-frame swap yields usable (s, w*) targets for Cosmos-Policy, but unlike the action-head case (App. B.1, Table 5: Action MSE and downstream SR by inverter), the manuscript does not report reconstruction error or success-rate sensitivity for the WAM inverter (joint vs. action-frame-only, with/without terminal Adam). Table 2 shows task gains, which is necessary but not sufficient to establish that inversion fidelity—not other factors—drives those gains. A short quantitative reconstruction table for Cosmos-Policy analogous to Table 5 would close this gap.","section":null},{"comment":"§4.3 dual-buffer training (equal mix of inverted corrections and successful autonomous noises) is presented as essential to avoid overfitting sparse interventions, yet no ablation of the buffer mix or of the autonomous buffer alone appears in the main results. Given that free parameters listed for the method include this ratio, a small sensitivity study (e.g., corrections-only vs. 50/50 vs. autonomous-heavy) on one MetaWorld task would show whether the dual buffer is load-bearing for the reported sample efficiency.","section":null}],"minor_comments":[{"comment":"Fig. 3 caption and §5.1: clarify whether intervention frequency and human gating criteria are matched across FlowDAgger, Residual-DAgger, and LoRA-DAgger, or only the total rollout budget N=50.","section":null},{"comment":"Eq. (3) and App. B.2: state the practical stopping criterion or residual tolerance used online, not only M=5 as default, so implementers can verify contraction on new base models.","section":null},{"comment":"Table 3: held-out prior preservation is shown only for π0.5 after Hammer adaptation; a one-sentence note on whether similar checks were run for Cosmos-Policy or real tasks would help scope the claim in §5.4.","section":null},{"comment":"App. A.3: the PCA basis rank k=64 for WAM noise is a free parameter; briefly report how k was chosen and whether full joint regression was unstable only in wall-clock or also in SR.","section":null},{"comment":"Minor notation: π_gp(s, w) is used for the noise-to-action map while π_gp(a|s) denotes the induced distribution; a short glossary or consistent bolding of noise vs. action would reduce ambiguity in §3–4.","section":null},{"comment":"Website link is given; if code/inversion scripts will be released, stating that in the camera-ready would strengthen reproducibility claims.","section":null}],"recommendation":"minor_revision","confidential_remarks":"Solid empirical systems paper for a robotics venue. The real-hardware baseline gap and missing WAM inversion metrics are fixable without new theory; I would not block on them if the authors add Residual-DAgger on at least a subset of real tasks and a short WAM reconstruction table. Scope fits cs.RO / robot learning well. No novelty or citation concerns stood out."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The thing worth knowing is that they turn human corrective actions into noise targets for a frozen flow/diffusion policy, then train a tiny steerer on those targets plus successful autonomous noise. That is the operational novelty: per-step fixed-point inversion for few-step action heads (Eq. 3), plus a joint-latent variant for world-action models, feeding ordinary supervised regression rather than RL or weight updates.\n\nWhat works: the design is coherent and the comparisons are the right ones. On MetaWorld with π0.5 they beat SFT, LoRA-DAgger, Residual-DAgger, and DSRL under matched 50-rollout budgets (Table 1, Fig. 3), transfer the same recipe to Cosmos-Policy (Table 2), show better prior preservation on held-out tasks than weight-space methods (Table 3), and report real gains on eight hardware tasks from 5–20 interventions (Table 4). Appendix inversion ablations (B.1–B.2) and replications on Gr00t and a vanilla diffusion policy make the technical claim more than a single-base anecdote. Compute story is honest: only the small noise policy is trained, so it fits consumer GPU memory.\n\nSoft spots are real but proportionate. Real-robot numbers lack seeds and error bars, so treat Table 4 as directional. The dual-buffer ratio, M, and WAM PCA rank are free parameters; they are not hidden, just not exhaustively swept. The load-bearing bound is the one they already state in §6: if the desired correction sits far outside the base manifold, inversion recovers only the nearest representable behavior. All reported gains sit inside that regime (low Action MSE in App. B.1). No internal contradiction or circular metric.\n\nThis is for people adapting large generative robot policies under sparse human time and frozen-base constraints. It is not a theory paper. I would send it to peer review; the method is clear, the baselines are fair, and the limitations are named. Worth engaging if you work on foundation-model adaptation or human-in-the-loop robotics.","headline":"Clean, practical method for steering frozen generative robot policies from human corrections via action inversion; evidence is solid within the stated manifold-support limit.","tokens_in":15553,"tokens_out":529,"would_cite":true,"duration_ms":5461,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Human corrections can adapt frozen generative robot policies by mapping each expert action into the noise that would have produced it, then training a small latent steerer.","keywords":["generative robot policies","flow matching","diffusion policy","human-in-the-loop","DAgger","latent-space adaptation","action inversion","robotic manipulation"],"falsifier":"On a task where the expert correction is known to lie outside the base policy’s action manifold (for example a contact strategy never seen in pretraining), measure whether inverted noise still reconstructs the expert action closely and whether the adapted noise policy improves success; failure of both reconstruction and adaptation would falsify the method’s core claim for out-of-support skills.","tokens_in":15518,"feed_emoji":"🤖","tokens_out":639,"duration_ms":9218,"temperature":0.7,"pith_summary":"Pretrained flow-matching and diffusion robot policies encode strong behavioral priors, but fail on real-world edge cases outside their training mixture. Retraining or fine-tuning the whole model is slow, expensive, and often erases skills the base already had. FlowDAgger keeps the base frozen and instead converts each human corrective action into the noise vector that would have made the base policy output that action, using reverse-time integration and local refinement. Those inverted noises supervise a lightweight latent policy that chooses noise at deployment, so the same generative process is steered without changing its weights. Across simulation and real single-arm and bimanual hardware, including action-head VLAs and world-action models, a handful of interventions raises success while held-out pretrained skills stay largely intact.","feed_headline":"Few human fixes steer frozen robot policies via noise inversion","feed_subtitle":"Map expert actions to latent noise, train a tiny steerer, keep the base and its skills intact.","key_machinery":"Action inversion: reverse-time per-step fixed-point integration (and a joint world-action variant for world-action models) that maps each expert action a* at observation s to noise w* with π_gp(s, w*) ≈ a*, providing targets for a lightweight noise policy that replaces the usual Gaussian draw at deployment.","core_discovery":"The paper claims that action inversion—recovering a noise vector w* such that a frozen generative policy maps (observation s, w*) to a human corrective action a*—turns sparse human interventions into supervision for a small noise-space policy, enabling sample- and compute-efficient adaptation of flow-matching and diffusion robot policies without weight updates and with better preservation of pretrained skills than fine-tuning or action-space residual methods.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Action inversion maps human fixes to noise for frozen robot policy steering","Sparse interventions adapt generative robots via latent noise recovery","Invert expert actions to noise then train tiny steerer on frozen policies","Human corrections become inverted noise to guide frozen flow robot models","Recover latent noise from few fixes to adapt robots without weight updates"],"cache_read_input_tokens":128,"weakest_assumption_plain":"Desired corrections must already lie close enough to behaviors the frozen base can produce; if the needed skill is far outside that manifold, inversion only recovers the nearest representable base behavior and cannot invent new ones.","fun_headline_variants_meta":{"raw":{"variants":["Action inversion maps human fixes to noise for frozen robot policy steering","Sparse interventions adapt generative robots via latent noise recovery","Invert expert actions to noise then train tiny steerer on frozen policies","Human corrections become inverted noise to guide frozen flow robot models","Recover latent noise from few fixes to adapt robots without weight updates"]},"model":"grok-4.5","effort":"low","cost_usd":0.003826,"raw_usage":{"total_tokens":1231,"prompt_tokens":794,"num_sources_used":0,"completion_tokens":70,"cost_in_usd_ticks":38260000,"prompt_tokens_details":{"text_tokens":794,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":367,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":794,"tokens_out":70,"duration_ms":4331,"temperature":1.0,"reasoning_tokens":367,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T06:04:28.394527+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a task where the expert correction is known to lie outside the base policy’s action manifold (for example a contact strategy never seen in pretraining), measure whether inverted noise still reconstructs the expert action closely and whether the adapted noise policy improves success; failure of both reconstruction and adaptation would falsify the method’s core claim for out-of-support skills.","supporting_citations":[],"review_version":1}